AI dictionary: every AI term explained simplyAI dictionary: har AI shabd aasaan bhasha mein

Product and careerProduct aur career Also calledIse yeh bhi kehte hain: Split test, online experiment 🔊

🧸 Think of it likeAise samjho: Like a sweet shop selling the same laddoo in two box designs for a week, then keeping the box that sold more.Jaise mithai ki dukaan ek hafte tak same laddoo do alag box designs mein bechti hai, phir jo box zyada bika woh rakh leti hai.

Users are split at random into a control group (the current version) and a treatment group (the new version). Because the split is random, the only systematic difference between the groups is the change, so a difference in results can be called a real effect rather than a coincidence. You decide the main metric and the guardrail metrics before starting, run for the planned time, and check whether the difference is statistically significant.Users ko randomly control group (purana version) aur treatment group (naya version) mein baanta jaata hai. Split random hai, isliye groups ke beech systematic fark sirf badlaav ka hota hai, aur results ka fark sach mein asar maana ja sakta hai, sanyog nahi. Main metric aur guardrail metrics shuru se pehle tay karo, planned time tak chalao, aur check karo ki fark statistically significant hai ya nahi.

💡 Real exampleAsli example: Half of the users see AI reply suggestions, half don't; you compare how fast tickets get solved.Aadhe users ko AI reply suggestions dikhte hain, aadhon ko nahi; phir compare karte ho tickets kitni jaldi solve hue.

📖 Learn itSeekho: Metrics and A/B testing for AI featuresAI features ke liye metrics aur A/B testing · Correlation vs causation, and sampling biasCorrelation vs causation, aur sampling bias

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Guardrail metric · Metric

Deep learningDeep learning 🔊

🧸 Think of it likeAise samjho: Like cooking dal three times, each time leaving out one spice, to find out which spice really makes it tasty.Jaise dal teen baar banana, har baar ek masala chhod kar, yeh pata karne ke liye ki asli swaad kaunsa masala deta hai.

If a new model has three changes and scores better, an ablation study turns off each change in turn and measures the score again. It shows which changes matter, which add only complexity, and makes research results believable. Good ablations change one thing at a time and repeat runs with different random seeds so differences aren't noise.Agar naye model mein teen badlaav hain aur score behtar hai, to ablation study baari baari har badlaav band karke score dobara naapti hai. Isse pata chalta hai kaunse badlaav zaroori hain, kaunse sirf complexity badhaate hain, aur research ke results pe bharosa hota hai. Acchi ablations ek baar mein ek cheez badalti hain aur alag random seeds se runs dohraati hain taaki fark sirf noise na ho.

💡 Real exampleAsli example: Removing the reranker from a RAG system drops answer accuracy from 86% to 74%, so the reranker clearly helps.RAG system se reranker hataane pe answer accuracy 86% se 74% ho jaati hai, matlab reranker saaf madad kar raha hai.

📖 Learn itSeekho: Reading and reproducing research papersResearch papers padhna aur reproduce karna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Benchmark · Evals (evaluation)

Machine learningMachine learning 🔊

🧸 Think of it likeAise samjho: Like your exam score: 45 out of 50 questions right means 90%. But if almost every answer is 'No', always writing 'No' also scores high, so accuracy alone can fool you.Jaise exam ka score: 50 mein se 45 sahi matlab 90%. Par agar lagbhag har jawaab 'Nahi' hai, toh hamesha 'Nahi' likhne se bhi score achha aayega, isliye sirf accuracy dhokha de sakti hai.

Accuracy = correct predictions ÷ all predictions. It's easy to understand but misleading when one class is rare: if 1% of payments are fraud, a model that always says 'not fraud' is 99% accurate and completely useless. For imbalanced problems, look at precision, recall and the confusion matrix, and always compare against a simple baseline.Accuracy = sahi predictions ÷ saari predictions. Samajhna aasaan hai par tab misleading hai jab ek class rare ho: agar 1% payments fraud hain, to hamesha 'fraud nahi' bolne wala model 99% accurate hai aur bilkul bekaar. Imbalanced problems mein precision, recall aur confusion matrix dekho, aur hamesha simple baseline se compare karo.

💡 Real exampleAsli example: A spam filter that's right on 950 of 1,000 emails has 95% accuracy.Jo spam filter 1,000 mein se 950 emails pe sahi hai, uski accuracy 95% hai.

📖 Learn itSeekho: Measuring models: accuracy, precision and recallModels ko measure karna: accuracy, precision aur recall

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Precision · Recall · Confusion matrix · Baseline

Deep learningDeep learning Also calledIse yeh bhi kehte hain: ReLU, sigmoid, GELU 🔊

🧸 Think of it likeAise samjho: Like a doorkeeper for each brain cell: weak signals are stopped at the door, strong ones are let through. That on and off choice lets the network learn bends and curves.Har brain cell ke liye darbaan jaisa: kamzor signal darwaaze pe ruk jaate hain, strong andar jaate hain. Yahi on off ka faisla network ko mod aur curves seekhne deta hai.

Each neuron computes a weighted sum of its inputs and passes it through an activation function. Without it, stacking layers would still only produce a straight line relationship. ReLU (keep positives, zero negatives) is the common default in hidden layers; GELU is used in transformers; sigmoid and softmax turn outputs into probabilities.Har neuron apne inputs ka weighted sum nikaalta hai aur use activation function se guzaarta hai. Iske bina layers jodne pe bhi sirf seedhi line wala rishta banta. Hidden layers mein ReLU (positive rakho, negative ko zero) common default hai; transformers mein GELU; sigmoid aur softmax outputs ko probabilities mein badalte hain.

💡 Real exampleAsli example: ReLU turns -3 into 0 and keeps 5 as 5.ReLU -3 ko 0 bana deta hai aur 5 ko 5 hi rehne deta hai.

📖 Learn itSeekho: Neural networks: layers that learnNeural networks: layers jo seekhti hain

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Neural network · Softmax

Agents and toolsAgents aur tools Also calledIse yeh bhi kehte hain: Autonomous agent 🔊

🧸 Think of it likeAise samjho: Like a smart assistant you give a goal, not steps: 'book me a cheap train to Pune on Friday'. It searches, compares, checks your calendar and comes back to you before paying.Ek smart assistant jaisa jise aap goal dete ho, steps nahi: 'Friday ko Pune ki sasti train book karo'. Woh search karta hai, compare karta hai, calendar check karta hai aur pay karne se pehle aapse poochta hai.

An agent gets a goal and a set of tools (search, database lookup, sending an email). In each turn the model decides which tool to call; your code runs the tool and gives the result back. This suits open ended tasks where the steps aren't known in advance. Because each step costs time and money and can go wrong, agents need limits (max steps, budgets), human approval for risky actions, tracing and evaluation. If the steps are fixed, a simple workflow is better.Agent ko ek goal aur kuch tools milte hain (search, database lookup, email bhejna). Har turn mein model tay karta hai kaunsa tool call karna hai; aapka code tool chalata hai aur result wapas deta hai. Yeh un open ended kaamon ke liye accha hai jinke steps pehle se pata nahi. Har step mein time aur paisa lagta hai aur galti ho sakti hai, isliye agents ko limits (max steps, budget), risky actions pe human approval, tracing aur evaluation chahiye. Steps fixed hon to simple workflow behtar hai.

💡 Real exampleAsli example: A support agent looks up the order, checks the refund policy and drafts a reply for a human to approve.Support agent order dhoondhta hai, refund policy check karta hai aur insaan ke approval ke liye reply draft karta hai.

📖 Learn itSeekho: AI agents: plan, act, observe, repeatAI agents: plan, act, observe, repeat · Tool use (function calling): letting an LLM take actionsTool use (function calling): LLM ko action lene dena

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Tool calling · MCP (Model Context Protocol) · Guardrails · Workflow (AI workflow)

AI basicsAI basics 🔊

AI is the big umbrella. Older AI used hand written rules; most modern AI is machine learning, where the system learns patterns from examples. Deep learning and generative AI (like ChatGPT or Claude) are parts of machine learning. In jobs, 'AI' usually means building products and systems around these models.AI ek bada umbrella hai. Purana AI haath se likhe rules use karta tha; aaj ka zyaadatar AI machine learning hai, jahan system examples se patterns seekhta hai. Deep learning aur generative AI (jaise ChatGPT ya Claude) machine learning ke hisse hain. Jobs mein 'AI' ka matlab aksar in models ke around products aur systems banana hota hai.

💡 Real exampleAsli example: Face unlock, map traffic predictions and voice assistants are all AI.Face unlock, map ki traffic predictions aur voice assistants, sab AI hain.

📖 Learn itSeekho: What is AI? AI vs ML vs Deep Learning vs GenAIAI kya hai? AI vs ML vs Deep Learning vs GenAI

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Machine learning · Deep learning · GenAI (generative AI)

Product and careerProduct aur career Also calledIse yeh bhi kehte hain: LLM engineer, applied AI engineer 🔊

🧸 Think of it likeAise samjho: Like an electrician who doesn't make the power plant (the AI model) but wires it safely into your house so the lights and fans actually work.Ek electrician jaisa jo power plant (AI model) nahi banata, par use aapke ghar mein safely jodta hai taaki lights aur pankhe sach mein chalein.

AI engineers usually don't train big models from scratch. They connect models to real data and apps: prompts, RAG, tool calling, agents, evaluation, guardrails, latency and cost. Strong software engineering matters more than heavy maths. Related roles include ML engineer (trains and serves models), MLOps engineer (runs them reliably) and AI product engineer (ships user facing AI features).AI engineers aam taur pe bade models shuru se train nahi karte. Woh models ko asli data aur apps se jodte hain: prompts, RAG, tool calling, agents, evaluation, guardrails, latency aur cost. Bhaari maths se zyaada mazboot software engineering zaroori hai. Jude roles: ML engineer (models train aur serve karta hai), MLOps engineer (unhe reliably chalata hai) aur AI product engineer (users ke liye AI features ship karta hai).

💡 Real exampleAsli example: Building a company chatbot that answers from HR policies with citations is typical AI engineer work.HR policies se citations ke saath jawaab dene wala company chatbot banana typical AI engineer ka kaam hai.

📖 Learn itSeekho: Landing the AI role: resume, applications and interviewsAI role paana: resume, applications aur interviews · Calling an LLM API: keys, requests and costsLLM API call karna: keys, requests aur costs

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: RAG (retrieval augmented generation) · Agent (AI agent) · MLOps

Evaluation and safetyEvaluation aur safety 🔊

🧸 Think of it likeAise samjho: Like training a very capable new employee in the company's values, so they are helpful to customers, honest, and refuse harmful requests.Jaise ek bahut kaabil naye employee ko company ki values sikhaana, taaki woh customers ki madad kare, sach bole, aur nuksaan wali requests mana kar de.

A model trained only to predict text can produce harmful, false or unhelpful output. Alignment techniques, such as instruction tuning, RLHF, DPO and constitutional methods, teach the model to follow instructions, refuse harmful requests and admit uncertainty. In products, alignment is supported by system prompts, guardrails and evaluation.Sirf text predict karne ke liye train hua model nuksaandeh, galat ya bekaar output de sakta hai. Alignment techniques, jaise instruction tuning, RLHF, DPO aur constitutional methods, model ko instructions maanna, nuksaandeh requests mana karna aur anishchitata maanna sikhaati hain. Products mein alignment ko system prompts, guardrails aur evaluation sahaara dete hain.

💡 Real exampleAsli example: A chatbot politely refusing to explain how to make a weapon is alignment at work.Chatbot ka hathiyar banane ka tarika batane se vinamrata se mana karna alignment ka kaam hai.

📖 Learn itSeekho: AI safety in products: hallucinations, prompt injection and guardrailsProducts mein AI safety: hallucinations, prompt injection aur guardrails

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: RLHF · Guardrails · Hallucination

DataData Also calledIse yeh bhi kehte hain: Data labelling 🔊

🧸 Think of it likeAise samjho: Like a teacher writing the right answer next to each practice question, so the student can learn by checking.Jaise teacher har practice question ke saamne sahi jawaab likh deta hai, taaki student check karke seekh sake.

Supervised learning needs labelled examples: emails marked spam or not, images with boxes around cars, tickets tagged by category. Labels are made by people, by rules, or by an LLM checked by people. Clear labelling guidelines and checking agreement between labellers matter, because the model copies label mistakes.Supervised learning ko labelled examples chahiye: spam ya not spam mark kiye emails, gaadiyon ke around box wali images, category se tag kiye tickets. Labels log, rules, ya insaanon se check kiya LLM banata hai. Saaf labelling guidelines aur labellers ke beech agreement check karna zaroori hai, kyunki model label ki galtiyaan bhi copy karta hai.

💡 Real exampleAsli example: Tagging 2,000 support tickets as billing, delivery or technical to train a classifier.Classifier train karne ke liye 2,000 support tickets ko billing, delivery ya technical tag karna.

📖 Learn itSeekho: How machines learn: features, labels and trainingMachines kaise seekhti hain: features, labels aur training

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Label · Supervised learning · Dataset

AI basicsAI basics Also calledIse yeh bhi kehte hain: Application Programming Interface 🔊

🧸 Think of it likeAise samjho: Like a waiter in a restaurant: you don't go into the kitchen, you tell the waiter what you want and the waiter brings it back.Restaurant ke waiter jaisa: aap kitchen mein nahi jaate, waiter ko batate ho kya chahiye aur woh le aata hai.

An API defines what you can request and what you get back, usually as JSON over HTTP. LLM providers offer APIs: you send a model name and messages, and receive the generated text and token counts. In AI products you both call APIs (LLMs, databases) and expose your own (for example a /predict endpoint for your model).API tay karta hai aap kya maang sakte ho aur badle mein kya milega, aam taur pe HTTP pe JSON. LLM providers APIs dete hain: aap model ka naam aur messages bhejte ho, aur generated text aur token counts milte hain. AI products mein aap APIs call bhi karte ho (LLMs, databases) aur apni bhi dete ho (jaise model ke liye /predict endpoint).

💡 Real exampleAsli example: A weather app calls a weather API to get today's forecast.Weather app aaj ka forecast lene ke liye weather API call karta hai.

📖 Learn itSeekho: Calling an LLM API: keys, requests and costsLLM API call karna: keys, requests aur costs · Serving a model as an API with FastAPIModel ko FastAPI se API ke roop mein serve karna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Inference · Latency

Deep learningDeep learning 🔊

🧸 Think of it likeAise samjho: Like reading a sentence with a highlighter: to understand the word 'it', your eyes jump back to the words that tell you what 'it' is.Highlighter ke saath sentence padhne jaisa: 'it' samajhne ke liye aapki nazar wapas un shabdon pe jaati hai jo batate hain ki 'it' kya hai.

Each token creates a query, a key and a value vector. Comparing a token's query with every key gives attention weights (how relevant each other token is), and the token's new representation is a weighted mix of the values. So 'bank' near 'river' ends up different from 'bank' near 'loan'. Attention cost grows with the square of the text length, which is why long contexts are expensive.Har token ek query, ek key aur ek value vector banata hai. Token ki query ko har key se compare karne pe attention weights milte hain (baaki token kitne relevant hain), aur token ki nayi representation values ka weighted mix hoti hai. Isliye 'river' ke paas wala 'bank' 'loan' ke paas wale 'bank' se alag ban jaata hai. Attention ka kharch text ki lambai ke square ke saath badhta hai, isliye lambe contexts mehenge hain.

💡 Real exampleAsli example: In 'The animal didn't cross the road because it was tired', attention links 'it' to 'animal'.'The animal didn't cross the road because it was tired' mein attention 'it' ko 'animal' se jodta hai.

📖 Learn itSeekho: Transformers and attention: how LLMs readTransformers aur attention: LLMs kaise padhte hain

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Transformer · KV cache · Token

Machine learningMachine learning Also calledIse yeh bhi kehte hain: Area under the curve 🔊

🧸 Think of it likeAise samjho: Like checking how well a guard can tell thieves from shoppers: pick one thief and one shopper at random, and AUC is how often the guard rates the thief as more suspicious.Jaise check karna ki guard chor aur customer mein kitna farak kar paata hai: ek chor aur ek customer random chuno, AUC batata hai ki kitni baar guard chor ko zyada shaky maanta hai.

ROC AUC is the chance that the model scores a random positive higher than a random negative: 0.5 is guessing, 1.0 is perfect. PR AUC summarises precision and recall together and is more informative when positives are rare, like fraud. AUC doesn't pick a threshold for you; you still choose one for the business trade off.ROC AUC woh chance hai ki model ek random positive ko random negative se zyaada score de: 0.5 matlab andaaza, 1.0 matlab perfect. PR AUC precision aur recall ko saath mein summarise karta hai aur tab zyaada kaam ka hai jab positives rare hon, jaise fraud. AUC aapke liye threshold nahi chunta; business trade off ke liye threshold aapko khud chunna hai.

💡 Real exampleAsli example: A fraud model with ROC AUC 0.92 ranks fraud above normal payments 92% of the time.ROC AUC 0.92 wala fraud model 92% baar fraud ko normal payments se upar rank karta hai.

📖 Learn itSeekho: Measuring models: accuracy, precision and recallModels ko measure karna: accuracy, precision aur recall

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Precision · Recall · Threshold

MLOps and deploymentMLOps aur deployment Also calledIse yeh bhi kehte hain: HPA 🔊

🧸 Think of it likeAise samjho: Like a supermarket opening more billing counters when the queue gets long and closing them when it's quiet.Jaise supermarket line lambi hone pe zyada billing counters khol deta hai aur khaali hone pe band kar deta hai.

When load rises, more copies (replicas) of the service start; when it falls, they're removed to save cost. Kubernetes' Horizontal Pod Autoscaler scales on CPU, memory or custom metrics; for LLM servers, queue length or tokens per second are better signals. Models take time to load, so plan headroom and readiness checks.Load badhne pe service ki aur copies (replicas) start hoti hain; ghatne pe cost bachaane ke liye hata di jaati hain. Kubernetes ka Horizontal Pod Autoscaler CPU, memory ya custom metrics pe scale karta hai; LLM servers ke liye queue length ya tokens per second behtar signals hain. Models load hone mein time lete hain, isliye headroom aur readiness checks plan karo.

💡 Real exampleAsli example: During a festival sale the recommendation service grows from 4 to 20 replicas, then shrinks at night.Festival sale mein recommendation service 4 se 20 replicas ho jaati hai, phir raat ko ghat jaati hai.

📖 Learn itSeekho: Kubernetes: running containers at scaleKubernetes: containers ko bade scale pe chalaana · Serving LLMs efficiently: GPUs, batching and costLLMs ko efficiently serve karna: GPUs, batching aur cost

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Kubernetes · Latency · Inference

Deep learningDeep learning Also calledIse yeh bhi kehte hain: Backprop 🔊

🧸 Think of it likeAise samjho: Like a cricket coach watching a lost match on replay and working out how much each player's mistake cost, so each one knows what to fix.Jaise cricket coach haare hue match ka replay dekh kar pata karta hai ki har player ki galti ka kitna asar hua, taaki har koi jaane kya sudhaarna hai.

After a forward pass produces a prediction and a loss, backpropagation applies the chain rule from the output back to the first layer, computing the gradient of the loss for every weight. An optimiser like SGD or Adam then nudges the weights. Frameworks such as PyTorch do this automatically (autograd).Forward pass se prediction aur loss milne ke baad, backpropagation output se pehli layer tak chain rule lagata hai aur har weight ke liye loss ka gradient nikaalta hai. Phir SGD ya Adam jaisa optimiser weights ko thoda khiskaata hai. PyTorch jaise frameworks yeh apne aap karte hain (autograd).

💡 Real exampleAsli example: Like a teacher tracing which step in a long sum caused the wrong answer.Jaise teacher dhoondhe ki lambe sawaal ka kaunsa step galat jawaab ki wajah bana.

📖 Learn itSeekho: Gradients and gradient descent: how models improveGradients aur gradient descent: models behtar kaise hote hain · Neural networks: layers that learnNeural networks: layers jo seekhti hain

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Gradient · Gradient descent · Loss function

Machine learningMachine learning 🔊

🧸 Think of it likeAise samjho: Like checking a new diet against simply eating normally: if the fancy plan doesn't beat the simple one, it isn't worth it.Jaise naye diet plan ko normal khaane se compare karna: agar fancy plan simple waale se behtar nahi, toh woh bekaar hai.

Baselines include 'always predict the most common class', 'predict yesterday's value', a simple rule, or the current human process. Without a baseline, a score like 85% means nothing. Baselines also catch bugs: if a fancy model barely beats a trivial one, something may be wrong.Baselines mein 'hamesha sabse common class predict karo', 'kal ki value predict karo', ek simple rule, ya aaj ka insaani process hota hai. Baseline ke bina 85% jaisa score kuch nahi batata. Baselines bugs bhi pakadte hain: agar fancy model ek simple wale ko mushkil se harata hai, to kuch gadbad ho sakti hai.

💡 Real exampleAsli example: If 90% of loans are repaid, a model must beat 90% accuracy to be useful.Agar 90% loans chukaaye jaate hain, to model ko kaam ka hone ke liye 90% accuracy harani hogi.

📖 Learn itSeekho: Measuring models: accuracy, precision and recallModels ko measure karna: accuracy, precision aur recall · Two kinds of prediction: regression and classificationDo tarah ki prediction: regression aur classification

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Accuracy · Benchmark

MLOps and deploymentMLOps aur deployment Also calledIse yeh bhi kehte hain: Offline scoring 🔊

🧸 Think of it likeAise samjho: Like a teacher correcting all 40 notebooks in one go at night, instead of correcting each one the moment a student hands it in.Jaise teacher raat ko saari 40 copies ek saath check karta hai, har copy milte hi turant check karne ki jagah.

Batch inference scores large sets, for example every customer's churn risk each night, and stores the results for apps to read. It's simpler and cheaper than real time serving and fine when predictions don't need to be instant. Real time inference is needed when a decision must happen within a request, like blocking a fraudulent payment.Batch inference bade sets score karta hai, jaise har raat har customer ka churn risk, aur results store karta hai jo apps padhti hain. Yeh real time serving se simple aur sasta hai aur tab theek hai jab predictions turant na chahiye hon. Real time inference tab chahiye jab faisla request ke andar hi lena ho, jaise fraud payment rokna.

💡 Real exampleAsli example: A bank scores all credit card customers every night for next day offers.Bank har raat saare credit card customers ko agle din ke offers ke liye score karta hai.

📖 Learn itSeekho: Training vs inference: how ML runs in productionTraining vs inference: ML production mein kaise chalta hai

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Inference · Latency · Data pipeline

Deep learningDeep learning 🔊

🧸 Think of it likeAise samjho: Like a teacher correcting 32 copies, then giving the class one round of feedback, instead of giving feedback after every single copy.Jaise teacher 32 copies check karke class ko ek baar feedback deta hai, har ek copy ke baad feedback dene ki jagah.

Training updates weights after each mini batch rather than after every example or the whole dataset. Larger batches use the GPU efficiently but need more memory; smaller batches are noisier but can generalise well. In LLM serving, batching many user requests together raises throughput and lowers cost per request.Training har example ya poore dataset ke baad nahi, har mini batch ke baad weights update karti hai. Bade batches GPU ka accha use karte hain par zyaada memory chahiye; chhote batches mein zyaada noise hota hai par generalise acche se kar sakte hain. LLM serving mein bahut saari user requests saath batch karne se throughput badhta aur per request cost ghatta hai.

💡 Real exampleAsli example: Training with batch size 32 means the model looks at 32 images before each weight update.Batch size 32 se training matlab model har weight update se pehle 32 images dekhta hai.

📖 Learn itSeekho: Gradients and gradient descent: how models improveGradients aur gradient descent: models behtar kaise hote hain · Serving LLMs efficiently: GPUs, batching and costLLMs ko efficiently serve karna: GPUs, batching aur cost

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Epoch · Gradient descent · GPU

Evaluation and safetyEvaluation aur safety 🔊

🧸 Think of it likeAise samjho: Like a board exam: every student gets the same paper, so the scores can be compared fairly.Board exam jaisa: har student ko same paper milta hai, isliye scores ko fairly compare kar sakte hain.

Public benchmarks (for maths, coding, reasoning, retrieval) let teams compare models on the same questions. They're useful but limited: models can be tuned to them, and a high public score doesn't guarantee good results on your task. For products, build your own eval set from real cases.Public benchmarks (maths, coding, reasoning, retrieval ke liye) teams ko same sawaalon pe models compare karne dete hain. Kaam ke hain par seemit: models unke liye tune ho sakte hain, aur ooncha public score aapke task pe acche results ki guarantee nahi. Products ke liye asli cases se apna eval set banao.

💡 Real exampleAsli example: Comparing two embedding models on the same retrieval benchmark before choosing one.Ek chunne se pehle do embedding models ko same retrieval benchmark pe compare karna.

📖 Learn itSeekho: Evaluating AI: how to know it actually worksAI ko evaluate karna: kaise pata chale ki sach mein kaam karta hai · Reading and reproducing research papersResearch papers padhna aur reproduce karna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Evals (evaluation) · Test set

Evaluation and safetyEvaluation aur safety Also calledIse yeh bhi kehte hain: Fairness 🔊

🧸 Think of it likeAise samjho: Like a cricket selector who has only ever watched players from one city: without meaning to, they keep picking from that city.Jaise ek cricket selector jisne sirf ek hi shehar ke players dekhe hain: bina chaahe woh usi shehar se chunta rehta hai.

Models learn patterns in their data, including historical unfairness and gaps. A hiring model trained on past hires can repeat past discrimination; a face model trained mostly on one group performs worse on others. Check performance across groups, fix data coverage, avoid sensitive features where inappropriate, and keep humans accountable for high stakes decisions.Models apne data ke patterns seekhte hain, purana anyaay aur kamiyaan samet. Purane hires pe train hua hiring model purana bhedbhaav dohra sakta hai; zyaadatar ek group pe train hua face model doosron pe kharab karta hai. Alag groups pe performance check karo, data coverage theek karo, jahan sahi na ho wahan sensitive features se bacho, aur high stakes faislon ke liye insaanon ko zimmedaar rakho.

💡 Real exampleAsli example: A loan model approving far fewer applicants from one region with the same income is a bias warning sign.Same income ke bawajood ek region ke kaafi kam applicants approve karne wala loan model bias ka warning sign hai.

📖 Learn itSeekho: Correlation vs causation, and sampling biasCorrelation vs causation, aur sampling bias · AI safety in products: hallucinations, prompt injection and guardrailsProducts mein AI safety: hallucinations, prompt injection aur guardrails

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Sampling bias · Explainability

Machine learningMachine learning 🔊

🧸 Think of it likeAise samjho: Like studying for an exam: only reading the summary misses things (too simple), but mugging up last year's paper word by word fails on new questions (too complex). The sweet spot is understanding the ideas.Exam ki padhai jaisa: sirf summary padhne se cheezein chhoot jaati hain (bahut simple), par pichhle saal ka paper ratne se naye sawaal pe fail (bahut complex). Sahi jagah hai concepts samajhna.

High bias means underfitting: the model is too simple, so it's wrong on training and new data alike. High variance means overfitting: it fits training data closely but fails on new data. Increasing complexity lowers bias but raises variance. You find the sweet spot using validation data, regularisation and more data.High bias matlab underfitting: model bahut simple hai, training aur naye dono data pe galat. High variance matlab overfitting: training data pe kareeb fit, par naye data pe fail. Complexity badhaane se bias ghatta aur variance badhta hai. Sweet spot validation data, regularisation aur zyaada data se milta hai.

💡 Real exampleAsli example: A straight line through curved sales data has high bias; a 50 level decision tree has high variance.Curved sales data pe seedhi line high bias hai; 50 level ka decision tree high variance.

📖 Learn itSeekho: Train/test split and overfittingTrain/test split aur overfitting

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Overfitting · Underfitting · Regularisation

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: Lexical search 🔊

🧸 Think of it likeAise samjho: Like the index at the back of a textbook: it finds pages that contain the exact word you look up.Textbook ke peeche wale index jaisa: woh pages dhoondhta hai jismein wahi exact shabd hai jo aap dhoondh rahe ho.

BM25 scores a document higher when it contains the query's words, especially rare words, and adjusts for document length. It's excellent for exact terms like product codes, names and error numbers, where embeddings are weak. Combining BM25 with vector search (hybrid search) usually improves RAG retrieval.BM25 document ko tab zyaada score deta hai jab usme query ke shabd, khaaskar rare shabd, hon, aur document ki lambai ke hisaab se adjust karta hai. Product codes, naam aur error numbers jaise exact terms ke liye badhiya hai, jahan embeddings kamzor hain. BM25 ko vector search ke saath milana (hybrid search) aksar RAG retrieval sudhaarta hai.

💡 Real exampleAsli example: Searching 'ERR_4012' finds the exact troubleshooting page with BM25.'ERR_4012' search karne pe BM25 exact troubleshooting page dhoondh leta hai.

📖 Learn itSeekho: RAG: letting an LLM answer from your documentsRAG: LLM ko aapke documents se jawaab dene dena · Embeddings and vector databases: searching by meaningEmbeddings aur vector databases: matlab se search karna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Hybrid search · Semantic search · RAG (retrieval augmented generation)

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: CoT, reasoning 🔊

🧸 Think of it likeAise samjho: Like a maths teacher saying 'show your working': writing each step makes careless mistakes less likely.Jaise maths teacher kehta hai 'steps dikhao': har step likhne se laaparwaahi wali galtiyaan kam hoti hain.

Writing out intermediate steps helps models on maths, logic and multi step problems. Many modern 'reasoning' models do this internally before answering. In products, you usually show users only the final answer, and you still check results, because reasoning that sounds right can be wrong.Beech ke steps likhne se models ko maths, logic aur multi step problems mein madad milti hai. Kai naye 'reasoning' models jawaab se pehle yeh andar hi karte hain. Products mein users ko aam taur pe sirf final jawaab dikhaate ho, aur results phir bhi check karte ho, kyunki sahi lagne wali reasoning bhi galat ho sakti hai.

💡 Real exampleAsli example: 'First find the total cost, then the discount, then the final price' improves a model's maths answer.'Pehle total cost nikalo, phir discount, phir final price' se model ka maths jawaab sudharta hai.

📖 Learn itSeekho: Prompting: giving clear instructions to an LLMPrompting: LLM ko saaf instructions dena

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Prompt engineering · Few shot prompting

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: Conversational AI 🔊

Old chatbots followed scripted rules and buttons. LLM chatbots understand free text and generate answers, often grounded in company documents with RAG and connected to tools for actions. Good chatbots show sources, say when they don't know, and hand over to a human when needed.Purane chatbots scripted rules aur buttons follow karte the. LLM chatbots khula text samajhte aur jawaab banate hain, aksar RAG se company documents pe based aur actions ke liye tools se jude. Acche chatbots sources dikhate hain, na pata ho to batate hain, aur zaroorat ho to insaan ko saunp dete hain.

💡 Real exampleAsli example: A bank's assistant that answers card questions from policy documents.Bank ka assistant jo policy documents se card ke sawaalon ka jawaab deta hai.

📖 Learn itSeekho: What is an LLM? Tokens, context and temperatureLLM kya hai? Tokens, context aur temperature · RAG: letting an LLM answer from your documentsRAG: LLM ko aapke documents se jawaab dene dena · UX for AI features: designing for uncertaintyAI features ke liye UX: uncertainty ke hisaab se design

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: LLM (large language model) · RAG (retrieval augmented generation) · Agent (AI agent)

LLMs and GenAILLMs aur GenAI 🔊

🧸 Think of it likeAise samjho: Like cutting a big pizza into slices so you can pick up exactly the piece you want.Badi pizza ko slices mein kaatne jaisa, taaki aap wahi tukda utha sako jo chahiye.

Models and vector search work best on focused passages. Chunks should hold a complete idea: split along headings, paragraphs and tables rather than every N characters, often a few hundred tokens with a small overlap, and store title, section, date and permissions as metadata. Test chunk sizes against real questions; bad chunking is a top cause of wrong RAG answers.Models aur vector search focused passages pe sabse accha kaam karte hain. Chunk mein poora idea hona chahiye: har N characters ke bajaye headings, paragraphs aur tables ke hisaab se todo, aksar kuch sau tokens thode overlap ke saath, aur title, section, date aur permissions metadata mein rakho. Asli sawaalon pe chunk sizes test karo; galat chunking galat RAG jawaabon ki badi wajah hai.

💡 Real exampleAsli example: A 40 page policy PDF becomes 120 chunks, one per section or paragraph group.40 page ki policy PDF 120 chunks ban jaati hai, har section ya paragraph group ka ek.

📖 Learn itSeekho: RAG: letting an LLM answer from your documentsRAG: LLM ko aapke documents se jawaab dene dena · Embeddings and vector databases: searching by meaningEmbeddings aur vector databases: matlab se search karna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: RAG (retrieval augmented generation) · Embedding · Vector database

MLOps and deploymentMLOps aur deployment Also calledIse yeh bhi kehte hain: Continuous integration and delivery 🔊

🧸 Think of it likeAise samjho: Like a factory quality check on every product before it leaves: each change is tested automatically, and only good ones are shipped.Factory mein har product nikalne se pehle quality check jaisa: har badlaav automatically test hota hai, aur sirf sahi wale bheje jaate hain.

Continuous integration runs tests on every change; continuous delivery deploys passing changes safely. For ML, add data validation, a small training run, evaluation against a fixed set before promotion, and gradual rollouts with rollback. Tools include GitHub Actions, Jenkins and GitLab CI.Continuous integration har badlaav pe tests chalata hai; continuous delivery paas hue badlaav safely deploy karta hai. ML mein data validation, chhota training run, promotion se pehle fixed set pe evaluation, aur rollback ke saath dheere rollouts jodo. Tools: GitHub Actions, Jenkins aur GitLab CI.

💡 Real exampleAsli example: A pull request automatically runs unit tests and an eval suite before it can be merged.Pull request merge hone se pehle apne aap unit tests aur eval suite chalata hai.

📖 Learn itSeekho: MLOps: experiment tracking and model registriesMLOps: experiment tracking aur model registries · Git and GitHub: saving and sharing your workGit aur GitHub: apna kaam save aur share karna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: MLOps · Model registry · Evals (evaluation)

Evaluation and safetyEvaluation aur safety Also calledIse yeh bhi kehte hain: Source attribution 🔊

🧸 Think of it likeAise samjho: Like the footnotes in a school project that say where each fact came from, so the teacher can check it.School project ke footnotes jaisa jo batate hain har fact kahan se aaya, taaki teacher check kar sake.

In RAG systems the prompt asks the model to cite the chunks it used, and the app links them. Citations build trust, make errors easy to spot, and can be checked automatically (does the quoted text exist in the source?).RAG systems mein prompt model se use kiye chunks cite karne ko kehta hai, aur app unhe link karti hai. Citations bharosa banaate hain, galtiyaan aasaani se dikhti hain, aur automatically check ho sakte hain (kya quote kiya text source mein hai?).

💡 Real exampleAsli example: 'You get 12 casual leaves a year [Leave Policy 2026, section 3.2].''Aapko saal mein 12 casual leaves milti hain [Leave Policy 2026, section 3.2].'

📖 Learn itSeekho: RAG: letting an LLM answer from your documentsRAG: LLM ko aapke documents se jawaab dene dena · UX for AI features: designing for uncertaintyAI features ke liye UX: uncertainty ke hisaab se design

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: RAG (retrieval augmented generation) · Grounding · Hallucination

Machine learningMachine learning Also calledIse yeh bhi kehte hain: Classifier 🔊

🧸 Think of it likeAise samjho: Like sorting the post into boxes: bills, letters, ads. Each item goes into one category.Daak ko dabbon mein chhaantne jaisa: bills, chitthiyaan, ads. Har cheez ek category mein jaati hai.

Classification models output a class or class probabilities. Binary classification has two classes; multi class has several. Common models are logistic regression, gradient boosted trees and neural networks; LLMs can also classify text with a good prompt. Evaluate with precision, recall, F1 and a confusion matrix.Classification models ek class ya class probabilities dete hain. Binary classification mein do classes; multi class mein kai. Common models: logistic regression, gradient boosted trees aur neural networks; accha prompt ho to LLMs bhi text classify kar sakte hain. Precision, recall, F1 aur confusion matrix se evaluate karo.

💡 Real exampleAsli example: Predicting whether a loan applicant will default (yes or no).Predict karna ki loan lene wala default karega ya nahi (haan ya nahi).

📖 Learn itSeekho: Two kinds of prediction: regression and classificationDo tarah ki prediction: regression aur classification · Measuring models: accuracy, precision and recallModels ko measure karna: accuracy, precision aur recall

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Regression · Precision · Recall

Machine learningMachine learning Also calledIse yeh bhi kehte hain: K means 🔊

🧸 Think of it likeAise samjho: Like tipping a box of mixed buttons on the table and grouping them by look, without anyone telling you the groups.Jaise mixed buttons ka dabba table pe ulat kar unhe dikhne ke hisaab se groups mein rakhna, bina kisi ke groups bataye.

Clustering is unsupervised learning: the algorithm finds natural groups in data, for example customer segments by behaviour. K means assigns each item to the nearest of K centres and moves the centres until stable. Results need human interpretation: the algorithm doesn't know what the groups mean.Clustering unsupervised learning hai: algorithm data mein natural groups dhoondhta hai, jaise behaviour ke hisaab se customer segments. K means har item ko K centres mein se sabse paas wale ko deta hai aur centres ko stable hone tak khiskaata hai. Results ko insaani samajh chahiye: algorithm nahi jaanta groups ka matlab kya hai.

💡 Real exampleAsli example: An online store discovers 'weekend bargain hunters' and 'premium regulars' groups.Online store 'weekend bargain hunters' aur 'premium regulars' groups dhoondhta hai.

📖 Learn itSeekho: How machines learn: features, labels and trainingMachines kaise seekhti hain: features, labels aur training

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Unsupervised learning · Embedding

Deep learningDeep learning Also calledIse yeh bhi kehte hain: CV 🔊

🧸 Think of it likeAise samjho: Giving a computer eyes: like how you glance at a photo and instantly see a dog, a car and a street.Computer ko aankhein dena: jaise aap photo dekhte hi turant kutta, gaadi aur sadak pehchaan lete ho.

Tasks include classification (what's in the image), detection (where objects are), segmentation (which pixels belong to what) and OCR (reading text). Modern vision uses convolutional networks and vision transformers, and multimodal LLMs can describe and reason about images.Tasks mein classification (image mein kya hai), detection (objects kahan hain), segmentation (kaunse pixels kiske hain) aur OCR (text padhna) hain. Aaj ka vision convolutional networks aur vision transformers use karta hai, aur multimodal LLMs images ka varnan aur un pe reasoning kar sakte hain.

💡 Real exampleAsli example: A phone camera reading a cheque amount, or a factory camera spotting defective parts.Phone camera ka cheque ka amount padhna, ya factory camera ka kharab parts pakadna.

📖 Learn itSeekho: Neural networks: layers that learnNeural networks: layers jo seekhti hain

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Multimodal · Deep learning

MLOps and deploymentMLOps aur deployment 🔊

🧸 Think of it likeAise samjho: Like a shopkeeper whose old rule 'people buy umbrellas in June' stops working because the monsoon now starts in July.Jaise dukaandaar ka purana rule 'June mein log chhaate khareedte hain' kaam karna band kar de kyunki ab monsoon July mein aata hai.

The inputs may look the same, but what they mean has changed, for example fraudsters change tactics or customer preferences shift after a price change. Detect it by tracking real performance when labels arrive, plus proxy signals. Fix it by retraining on recent data, sometimes with new features.Inputs same dikh sakte hain, par unka matlab badal gaya hai, jaise fraudsters tarike badalte hain ya price change ke baad customers ki pasand badalti hai. Labels aane pe asli performance track karke aur proxy signals se pakdo. Haal ke data pe retrain karke theek karo, kabhi naye features ke saath.

💡 Real exampleAsli example: A fraud model trained before UPI scams became common misses the new pattern.UPI scams common hone se pehle train hua fraud model naya pattern miss kar deta hai.

📖 Learn itSeekho: Monitoring models: drift and silent failuresModels ki monitoring: drift aur chupke se hone wale failures

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Data drift · Monitoring

Machine learningMachine learning 🔊

🧸 Think of it likeAise samjho: Like a report card for a security guard: thieves caught, thieves missed, innocent people stopped, and innocent people let through.Security guard ka report card jaisa: kitne chor pakde, kitne chhoot gaye, kitne begunaah roke, aur kitne begunaah jaane diye.

For yes or no problems it counts true positives, false positives, false negatives and true negatives. Precision, recall and accuracy are all calculated from it. Looking at the matrix shows which mistakes the model makes, which matters more than a single score.Haan ya nahi wali problems mein yeh true positives, false positives, false negatives aur true negatives ginta hai. Precision, recall aur accuracy sab isi se nikalte hain. Matrix dekhne se pata chalta hai model kaunsi galtiyaan karta hai, jo ek score se zyaada zaroori hai.

💡 Real exampleAsli example: Out of 10 frauds, 8 caught (TP), 2 missed (FN); 12 normal payments wrongly flagged (FP).10 fraud mein se 8 pakde (TP), 2 chhoote (FN); 12 normal payments galti se flag hue (FP).

📖 Learn itSeekho: Measuring models: accuracy, precision and recallModels ko measure karna: accuracy, precision aur recall

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Precision · Recall · Accuracy

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: Context length 🔊

🧸 Think of it likeAise samjho: Like the size of a desk: everything the model can look at right now must fit on it. Older papers fall off the edge.Desk ke size jaisa: model abhi jo kuch dekh sakta hai woh sab us par aana chahiye. Purane kaagaz kinaare se gir jaate hain.

Everything the model sees, system prompt, documents, chat history, tool results and output, must fit in the context window. Bigger windows allow longer documents but cost more and are slower, and models can miss details in the middle of very long inputs. RAG and summarising history keep context focused.Model jo kuch dekhta hai, system prompt, documents, chat history, tool results aur output, sab context window mein fit hona chahiye. Badi windows lambe documents allow karti hain par mehngi aur slow hain, aur bahut lambe inputs ke beech ki details model miss kar sakta hai. RAG aur history ka summary context focused rakhte hain.

💡 Real exampleAsli example: A 200,000 token window holds roughly a 500 page book.200,000 token ki window lagbhag 500 page ki kitaab sambhaal leti hai.

📖 Learn itSeekho: What is an LLM? Tokens, context and temperatureLLM kya hai? Tokens, context aur temperature · Calling an LLM API: keys, requests and costsLLM API call karna: keys, requests aur costs

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Token · RAG (retrieval augmented generation) · Lost in the middle

LLMs and GenAILLMs aur GenAI 🔊

🧸 Think of it likeAise samjho: Like checking if two people are pointing in the same direction: it doesn't matter how long their arms are, only where they point.Jaise check karna ki do log ek hi disha mein ishaara kar rahe hain ya nahi: haath kitna lamba hai farak nahi padta, sirf disha.

It's the cosine of the angle between two vectors: 1 means same direction (very similar meaning), 0 unrelated. Vector databases use it (or the closely related dot product) to find the chunks closest in meaning to a question.Yeh do vectors ke beech ke angle ka cosine hai: 1 matlab same direction (bahut milta julta matlab), 0 matlab koi rishta nahi. Vector databases ise (ya jude hue dot product ko) use karke sawaal ke matlab ke sabse paas ke chunks dhoondhte hain.

💡 Real exampleAsli example: 'reset my password' and 'forgot login' have high cosine similarity.'reset my password' aur 'forgot login' ki cosine similarity zyaada hai.

📖 Learn itSeekho: Embeddings and vector databases: searching by meaningEmbeddings aur vector databases: matlab se search karna · Vectors and matrices: the language of AIVectors aur matrices: AI ki bhasha

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Embedding · Vector · Semantic search

Deep learningDeep learning Also calledIse yeh bhi kehte hain: Log loss 🔊

🧸 Think of it likeAise samjho: Like a quiz that gives a small minus for an unsure wrong answer but a huge minus when you were 100% sure and still wrong.Ek quiz jaisa jo unsure galat jawaab pe thoda minus deta hai, par 100% sure hokar galat hone pe bahut bada minus.

Cross entropy compares predicted probabilities with the true class. Predicting 0.9 for the right class gives a small loss; predicting 0.01 for it gives a large loss. LLMs are trained with cross entropy on the next token. Combined with softmax, its gradient is simply prediction minus truth.Cross entropy predicted probabilities ko sahi class se compare karta hai. Sahi class ko 0.9 dena chhota loss deta hai; use 0.01 dena bada loss. LLMs agle token pe cross entropy se train hote hain. Softmax ke saath iska gradient bas prediction minus sach hai.

💡 Real exampleAsli example: A cat image predicted as 95% dog gives a big cross entropy loss.Billi ki image ko 95% kutta predict karna bada cross entropy loss deta hai.

📖 Learn itSeekho: Gradients and gradient descent: how models improveGradients aur gradient descent: models behtar kaise hote hain · Neural networks: layers that learnNeural networks: layers jo seekhti hain

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Loss function · Softmax

Machine learningMachine learning Also calledIse yeh bhi kehte hain: K fold 🔊

🧸 Think of it likeAise samjho: Like taking five different mock tests instead of one, so a single lucky or unlucky paper doesn't fool you.Ek ki jagah paanch alag mock tests dena, taaki ek lucky ya unlucky paper aapko dhokha na de.

In k fold cross validation, training data is split into k parts; the model trains on k minus 1 and validates on the remaining one, k times, and the scores are averaged. It's useful for small datasets and tuning. Use time based splits for time series, and group splits when one person has many rows.K fold cross validation mein training data k hisson mein baanta jaata hai; model k minus 1 pe train aur bache hue pe validate hota hai, k baar, aur scores ka average liya jaata hai. Chhote datasets aur tuning ke liye kaam ka. Time series ke liye time based splits, aur jab ek insaan ki kai rows hon to group splits.

💡 Real exampleAsli example: 5 fold cross validation gives scores 81, 84, 79, 83, 82; average 81.8.5 fold cross validation se scores 81, 84, 79, 83, 82; average 81.8.

📖 Learn itSeekho: Train/test split and overfittingTrain/test split aur overfitting

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Validation set · Overfitting

MLOps and deploymentMLOps aur deployment Also calledIse yeh bhi kehte hain: Covariate shift 🔊

🧸 Think of it likeAise samjho: Like a weather app trained on Delhi now being used in Chennai: the inputs look different from what it learned on.Jaise Delhi pe trained weather app ab Chennai mein use ho: inputs us data se alag dikhte hain jis par usne seekha.

New customer segments, new devices, a changed form or a broken upstream system can all shift input distributions. Monitor feature and prediction distributions against a baseline (for example with PSI), alert on big shifts, and investigate: sometimes it's a bug, sometimes the world changed and you need to retrain.Naye customer segments, naye devices, badla hua form ya toota upstream system, sab input distributions khiska sakte hain. Feature aur prediction distributions ko baseline se monitor karo (jaise PSI se), bade badlaav pe alert, aur jaanch karo: kabhi bug hota hai, kabhi duniya badal gayi hai aur retrain chahiye.

💡 Real exampleAsli example: After launching in a new city, average delivery distances double and the delivery time model drifts.Naye shehar mein launch ke baad average delivery distance double hoti hai aur delivery time model drift karta hai.

📖 Learn itSeekho: Monitoring models: drift and silent failuresModels ki monitoring: drift aur chupke se hone wale failures · Data quality: catching bad data earlyData quality: kharab data jaldi pakadna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Concept drift · Monitoring

Machine learningMachine learning Also calledIse yeh bhi kehte hain: Leakage 🔊

🧸 Think of it likeAise samjho: Like a student who saw the answer key before the exam: the score looks amazing, but it says nothing about what they really know.Jaise student ne exam se pehle answer key dekh li: score shaandaar dikhta hai, par asli samajh ke baare mein kuch nahi batata.

Examples: using 'refund issued' to predict fraud, scaling with statistics from the whole dataset before splitting, or the same customer in train and test. The model looks brilliant in testing and fails in production. Ask for every feature: would I know this at the moment of prediction?Examples: fraud predict karne ke liye 'refund issued' use karna, split se pehle poore dataset ke statistics se scaling, ya same customer train aur test dono mein. Model testing mein shaandaar dikhta hai aur production mein fail. Har feature ke liye poochho: prediction ke moment pe kya yeh mujhe pata hoga?

💡 Real exampleAsli example: A churn model scoring 0.99 AUC because it used 'account closed date' as a feature.'account closed date' ko feature banane se 0.99 AUC dene wala churn model.

📖 Learn itSeekho: Train/test split and overfittingTrain/test split aur overfitting · Feature engineering: turning raw data into useful inputsFeature engineering: raw data ko kaam ke inputs mein badalna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Feature · Test set · Overfitting

DataData Also calledIse yeh bhi kehte hain: ETL pipeline 🔊

🧸 Think of it likeAise samjho: Like a water supply system: water is collected, filtered and piped to every tap, automatically every day.Paani ki supply jaisa: paani ikattha hota hai, filter hota hai aur har nal tak pipe se pahunchta hai, roz automatically.

A pipeline extracts data (databases, APIs, files), loads and transforms it (SQL, dbt, Spark) and publishes clean tables, features or vector indexes. It's orchestrated on a schedule (Airflow) with retries, data quality checks and alerts. Good pipelines are idempotent, so reruns and backfills don't create duplicates.Pipeline data nikaalti hai (databases, APIs, files), load aur transform karti hai (SQL, dbt, Spark) aur saaf tables, features ya vector indexes publish karti hai. Schedule pe orchestrate hoti hai (Airflow), retries, data quality checks aur alerts ke saath. Acchi pipelines idempotent hoti hain, taaki reruns aur backfills duplicates na banayein.

💡 Real exampleAsli example: Every night, orders are copied from the app database, cleaned and turned into a sales dashboard table.Har raat orders app database se copy, saaf aur sales dashboard table mein badle jaate hain.

📖 Learn itSeekho: Data pipelines: ETL, Airflow, dbt and SparkData pipelines: ETL, Airflow, dbt aur Spark · Data quality: catching bad data earlyData quality: kharab data jaldi pakadna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: ETL and ELT · Idempotency · Feature store

DataData Also calledIse yeh bhi kehte hain: BigQuery, Snowflake, Redshift 🔊

🧸 Think of it likeAise samjho: Like a big, well organised godown of a company's history, arranged so you can quickly answer questions like 'sales by city last year'.Company ke itihaas ka ek bada, saaf suthra godown jaisa, aise lagaaya ki 'pichhle saal shehar ke hisaab se sales' jaise sawaal jaldi answer ho jaayein.

Warehouses store data in columnar form so analytics queries over millions of rows are fast. Teams load raw data into them (ELT), model it into fact and dimension tables, and use it for dashboards, metrics and ML training data. Lakehouses (like Databricks) combine warehouse features with cheap file storage.Warehouses data ko columnar roop mein rakhte hain taaki laakhon rows pe analytics queries tez chalein. Teams raw data load karti hain (ELT), use fact aur dimension tables mein model karti hain, aur dashboards, metrics aur ML training data ke liye use karti hain. Lakehouses (jaise Databricks) warehouse ke features ko saste file storage ke saath jodte hain.

💡 Real exampleAsli example: Running 'monthly revenue by city' over 3 years of orders in seconds.3 saal ke orders pe 'shehar ke hisaab se monthly revenue' seconds mein chalana.

📖 Learn itSeekho: Data modelling: designing tables for analysisData modelling: analysis ke liye tables design karna · Data pipelines: ETL, Airflow, dbt and SparkData pipelines: ETL, Airflow, dbt aur Spark

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: ETL and ELT · Data pipeline

DataData 🔊

🧸 Think of it likeAise samjho: Like a big recipe book of past examples that the model studies from.Pichhle examples ki ek badi recipe book jaisa, jisse model padhta hai.

A dataset might be rows in a table, images with labels, or question and answer pairs. Its quality, coverage and size often matter more than the model choice. Split it into training, validation and test sets, document where it came from, and check for duplicates, bias and leakage.Dataset table ki rows, labels wali images, ya sawaal jawaab ke jode ho sakte hain. Iski quality, coverage aur size aksar model ke chunaav se zyaada matter karte hain. Ise training, validation aur test sets mein baanto, kahan se aaya document karo, aur duplicates, bias aur leakage check karo.

💡 Real exampleAsli example: 50,000 past loan applications with whether each was repaid.50,000 purani loan applications, har ek chukaaya gaya ya nahi uske saath.

📖 Learn itSeekho: How machines learn: features, labels and trainingMachines kaise seekhti hain: features, labels aur training · Data quality: catching bad data earlyData quality: kharab data jaldi pakadna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Label · Test set · Annotation (labelling)

Machine learningMachine learning 🔊

🧸 Think of it likeAise samjho: Like the game '20 questions': each yes or no question narrows things down until you reach the answer.'20 questions' game jaisa: har haan ya na wala sawaal options kam karta hai jab tak jawaab na mil jaaye.

Each split picks the question that best separates the outcomes, like 'income above ₹50k?'. Trees are easy to explain but a single deep tree overfits. Random forests and gradient boosting combine many trees and are among the strongest models for tabular data.Har split woh sawaal chunta hai jo results ko sabse acche se alag kare, jaise 'income ₹50k se zyaada?'. Trees samjhaane mein aasaan hain par ek gehra tree overfit karta hai. Random forests aur gradient boosting kai trees jodte hain aur tabular data ke liye sabse mazboot models mein hain.

💡 Real exampleAsli example: Has the customer missed a payment? If yes, high risk; if no, check income next.Kya customer ne payment miss kiya? Haan to high risk; nahi to agla income check.

📖 Learn itSeekho: Decision trees, random forests and boostingDecision trees, random forests aur boosting

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Random forest · Gradient boosting

Deep learningDeep learning 🔊

🧸 Think of it likeAise samjho: Like learning to recognise a face in stages: first edges, then eyes and nose, then the whole face. Each layer builds on the one before.Chehra pehchaanna stages mein seekhne jaisa: pehle edges, phir aankh aur naak, phir poora chehra. Har layer pichhli wali pe banti hai.

Deep networks learn features directly from raw data: early layers detect simple patterns, later layers combine them into concepts. They need lots of data and compute (GPUs) and power modern vision, speech recognition and LLMs. For most work you start from pretrained models instead of training from scratch.Deep networks raw data se seedhe features seekhte hain: shuru ki layers simple patterns pakadti hain, baad ki unhe concepts mein jodti hain. Inhe bahut data aur compute (GPUs) chahiye aur yeh aaj ke vision, speech recognition aur LLMs chalaate hain. Zyaadatar kaam mein shuru se train karne ke bajaye pretrained models se shuru karte ho.

💡 Real exampleAsli example: Speech to text on your phone uses deep learning.Phone pe speech to text deep learning use karta hai.

📖 Learn itSeekho: Neural networks: layers that learnNeural networks: layers jo seekhti hain · What is AI? AI vs ML vs Deep Learning vs GenAIAI kya hai? AI vs ML vs Deep Learning vs GenAI

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Neural network · Transformer · GPU

MLOps and deploymentMLOps aur deployment Also calledIse yeh bhi kehte hain: Serving, productionising 🔊

🧸 Think of it likeAise samjho: Like moving a dish from your home kitchen to a restaurant menu: now real customers can order it any time.Jaise ek dish ko ghar ki kitchen se restaurant ke menu pe laana: ab asli customers kabhi bhi order kar sakte hain.

Deployment packages the model with its preprocessing (often in Docker), serves it behind an API or runs it in batches, and adds monitoring, scaling and rollback. Safe rollouts use shadow mode or canaries. Most of the real effort in AI projects happens here, after the notebook.Deployment model ko uske preprocessing ke saath pack karta hai (aksar Docker mein), API ke peeche serve karta ya batches mein chalata hai, aur monitoring, scaling aur rollback jodta hai. Safe rollouts shadow mode ya canaries use karte hain. AI projects ki zyaadatar asli mehnat notebook ke baad yahin hoti hai.

💡 Real exampleAsli example: Wrapping a fraud model in FastAPI, packaging it with Docker and running it on Kubernetes.Fraud model ko FastAPI mein wrap karna, Docker se pack karna aur Kubernetes pe chalana.

📖 Learn itSeekho: Serving a model as an API with FastAPIModel ko FastAPI se API ke roop mein serve karna · Docker: packaging your app so it runs anywhereDocker: app ko pack karna taaki kahin bhi chale · Training vs inference: how ML runs in productionTraining vs inference: ML production mein kaise chalta hai

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Inference · Docker · MLOps

Deep learningDeep learning Also calledIse yeh bhi kehte hain: Image generation 🔊

🧸 Think of it likeAise samjho: Like a sculptor starting with a rough block and chipping away a little at a time until a statue appears. Here the block is random noise.Jaise sculptor ek khurdure pathar se shuru karke thoda thoda chhilta hai jab tak murti na ban jaaye. Yahan pathar random noise hai.

During training the model learns to remove noise from images. To generate, it starts from random noise and repeatedly denoises, guided by a text prompt, until an image appears. Diffusion models power most AI image and some video generators.Training mein model images se noise hataana seekhta hai. Banane ke liye random noise se shuru karke, text prompt ke guidance mein, baar baar noise hataata hai jab tak image na ban jaaye. Zyaadatar AI image aur kuch video generators diffusion models se chalte hain.

💡 Real exampleAsli example: Typing 'a tiger in a Mumbai local train, watercolour' and getting a painting.'a tiger in a Mumbai local train, watercolour' likh kar painting paana.

📖 Learn itSeekho: What is AI? AI vs ML vs Deep Learning vs GenAIAI kya hai? AI vs ML vs Deep Learning vs GenAI

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: GenAI (generative AI) · Multimodal

MLOps and deploymentMLOps aur deployment Also calledIse yeh bhi kehte hain: Container 🔊

🧸 Think of it likeAise samjho: Like a tiffin box that carries the whole meal with its own spoon and chutney, so it tastes the same in any office.Tiffin box jaisa jo poora khaana apne chammach aur chutney ke saath le jaata hai, taaki kisi bhi office mein same swaad aaye.

A Dockerfile lists the base image, dependencies, code and start command; building it creates an image, and running it creates a container. This removes 'works on my machine' problems and is the standard way to ship ML services to servers and Kubernetes.Dockerfile mein base image, dependencies, code aur start command likhe hote hain; build karne se image banti hai, aur chalaane se container. Isse 'mere machine pe chalta hai' wali problems khatam hoti hain, aur ML services ko servers aur Kubernetes tak bhejne ka yahi standard tarika hai.

💡 Real exampleAsli example: Shipping your FastAPI model service as a Docker image to the cloud.Apni FastAPI model service ko Docker image ke roop mein cloud pe bhejna.

📖 Learn itSeekho: Docker: packaging your app so it runs anywhereDocker: app ko pack karna taaki kahin bhi chale

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Kubernetes · Deployment

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: Direct preference optimisation 🔊

🧸 Think of it likeAise samjho: Like showing a trainee two replies and saying 'this one is better, that one is worse', again and again, until they naturally write the better kind.Jaise trainee ko do replies dikha kar baar baar kehna 'yeh behtar hai, woh kharaab', jab tak woh khud behtar wala likhna na seekh le.

Given a prompt with a better and a worse response, DPO adjusts the model to make the preferred response more likely, without training a separate reward model. It's widely used to align open models with human preferences after instruction tuning.Ek prompt ke behtar aur badtar response diye hon to DPO model ko adjust karta hai taaki pasand wala response zyaada likely ho, bina alag reward model train kiye. Instruction tuning ke baad open models ko insaani pasand se align karne mein kaafi use hota hai.

💡 Real exampleAsli example: Training a support model to prefer short, polite, cited answers over long vague ones.Support model ko lambe vague jawaabon ke bajaye chhote, vinamra, cited jawaab pasand karna sikhaana.

📖 Learn itSeekho: Fine-tuning and LoRA: specialising a modelFine-tuning aur LoRA: model ko specialise karna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: RLHF · Fine tuning · Alignment

Deep learningDeep learning 🔊

🧸 Think of it likeAise samjho: Like a team practising with random players sitting out, so everyone learns every role and nobody depends on one star.Jaise team random players ko bitha kar practice kare, taaki sab har role seekhein aur koi ek star pe depend na rahe.

Dropout is a regularisation method: with, say, 10% probability each neuron's output is set to zero in each training step. This forces redundant, robust patterns and reduces overfitting. It's turned off when the model is used for predictions.Dropout regularisation ka tarika hai: maan lo 10% probability se har training step mein har neuron ka output zero kar diya jaata hai. Isse network mazboot, dohraaye patterns seekhta hai aur overfitting kam hoti hai. Predictions ke time ise band kar dete hain.

💡 Real exampleAsli example: Like a team practising with random players sitting out, so everyone learns every role.Jaise team random khiladiyon ko bitha kar practice kare, taaki sab har role seekhein.

📖 Learn itSeekho: Neural networks: layers that learnNeural networks: layers jo seekhti hain · Train/test split and overfittingTrain/test split aur overfitting

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Regularisation · Overfitting

Machine learningMachine learning 🔊

🧸 Think of it likeAise samjho: Like taking the cake out of the oven when it's perfectly done, instead of baking longer and burning it.Cake ko perfect banne pe oven se nikaal lena, zyada der bake karke jalaane ki jagah.

As training continues, training loss keeps falling but validation loss eventually rises as the model starts memorising. Early stopping watches validation performance and keeps the best checkpoint. It's standard for gradient boosting and neural networks.Training chalti rahe to training loss girta rehta hai par aakhir mein validation loss badhne lagta hai kyunki model ratne lagta hai. Early stopping validation performance dekhta hai aur best checkpoint rakhta hai. Gradient boosting aur neural networks mein standard hai.

💡 Real exampleAsli example: Validation loss is lowest at round 240, so training stops at 260 and keeps round 240.Validation loss round 240 pe sabse kam hai, to training 260 pe rukti hai aur round 240 rakhti hai.

📖 Learn itSeekho: Train/test split and overfittingTrain/test split aur overfitting · Decision trees, random forests and boostingDecision trees, random forests aur boosting

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Overfitting · Validation set · Epoch

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: Vector embedding 🔊

🧸 Think of it likeAise samjho: Like giving every word a location on a map, where words with similar meaning live in the same neighbourhood.Har shabd ko map pe ek pata dena, jahan milte julte matlab wale shabd ek hi mohalle mein rehte hain.

An embedding model turns a sentence into a vector of hundreds of numbers. Texts with similar meaning get nearby vectors even with different words. Embeddings power semantic search, RAG, recommendations, clustering and duplicate detection. Always use the same embedding model for documents and queries.Embedding model ek sentence ko sainkdon numbers ke vector mein badalta hai. Milte julte matlab wale texts ke vectors paas aate hain, shabd alag hon tab bhi. Embeddings semantic search, RAG, recommendations, clustering aur duplicates dhoondhna chalaati hain. Documents aur queries ke liye hamesha same embedding model use karo.

💡 Real exampleAsli example: 'Cheap flights to Goa' and 'low cost tickets Goa' get very close embeddings.'Cheap flights to Goa' aur 'low cost tickets Goa' ki embeddings bahut paas hoti hain.

📖 Learn itSeekho: Embeddings and vector databases: searching by meaningEmbeddings aur vector databases: matlab se search karna · RAG: letting an LLM answer from your documentsRAG: LLM ko aapke documents se jawaab dene dena · Vectors and matrices: the language of AIVectors aur matrices: AI ki bhasha

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Vector database · Cosine similarity · Semantic search

Deep learningDeep learning 🔊

🧸 Think of it likeAise samjho: Like reading your whole textbook once from cover to cover. Three epochs means three full readings.Poori textbook ek baar shuru se aakhir tak padhne jaisa. Teen epochs matlab teen baar poori padhai.

Training usually runs for several epochs, each made of many mini batches. Too few epochs underfit; too many overfit. Watch validation metrics per epoch and use early stopping. Large LLMs often train for about one epoch over enormous data.Training aam taur pe kai epochs chalti hai, har ek kai mini batches se bana. Bahut kam epochs underfit; bahut zyaada overfit. Har epoch pe validation metrics dekho aur early stopping use karo. Bade LLMs aksar bahut bade data pe lagbhag ek epoch train hote hain.

💡 Real exampleAsli example: With 10,000 images and batch size 100, one epoch is 100 steps.10,000 images aur batch size 100 ke saath ek epoch 100 steps ka hai.

📖 Learn itSeekho: Gradients and gradient descent: how models improveGradients aur gradient descent: models behtar kaise hote hain

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Batch size · Early stopping

DataData 🔊

🧸 Think of it likeAise samjho: Like groceries: ETL is washing and cutting vegetables before putting them in the fridge; ELT is putting everything in the fridge first and cutting what you need later.Sabzi jaisa: ETL matlab fridge mein rakhne se pehle dho kar kaat lena; ELT matlab sab pehle fridge mein rakhna aur zaroorat pe baad mein kaatna.

ETL cleans data before it lands, useful when data must be masked first. ELT loads raw data into a cloud warehouse and transforms it with SQL (often dbt), which is the common modern approach because raw data can be reprocessed when logic changes.ETL data aane se pehle saaf karta hai, jab data pehle mask karna zaroori ho tab kaam ka. ELT raw data cloud warehouse mein load karke SQL (aksar dbt) se transform karta hai, jo aaj ka common tarika hai kyunki logic badalne pe raw data dobara process ho sakta hai.

💡 Real exampleAsli example: Loading raw app events into BigQuery, then building clean tables with dbt.Raw app events BigQuery mein load karna, phir dbt se saaf tables banana.

📖 Learn itSeekho: Data pipelines: ETL, Airflow, dbt and SparkData pipelines: ETL, Airflow, dbt aur Spark

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Data pipeline · Data warehouse

Evaluation and safetyEvaluation aur safety Also calledIse yeh bhi kehte hain: Eval set, eval suite 🔊

🧸 Think of it likeAise samjho: Like a fixed set of practice tests you rerun after every change, to see if the system got better or worse.Practice tests ka ek fixed set jo har badlaav ke baad dobara chalaate ho, yeh dekhne ke liye ki system behtar hua ya kharaab.

For LLM apps, define what good means, collect real examples (common, edge cases, past failures), and score them with code checks where possible and rubrics (human or LLM judge) where needed. Run evals on every prompt or model change, like unit tests, and add each new production failure. Evals are what separate demos from reliable products.LLM apps ke liye accha ka matlab tay karo, asli examples ikattha karo (common, edge cases, purane failures), aur jahan ho sake code checks se, jahan zaroori rubrics (insaan ya LLM judge) se score karo. Har prompt ya model change pe unit tests ki tarah evals chalao, aur har naya production failure jodo. Evals hi demos ko reliable products se alag karte hain.

💡 Real exampleAsli example: 100 real customer questions with expected answers, rerun after every prompt change.Expected jawaabon ke saath 100 asli customer sawaal, har prompt change ke baad dobara chalaaye.

📖 Learn itSeekho: Evaluating AI: how to know it actually worksAI ko evaluate karna: kaise pata chale ki sach mein kaam karta hai · Writing an AI product spec (PRD) with an eval planEval plan ke saath AI product spec (PRD) likhna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: LLM as judge · Benchmark · Hallucination

Evaluation and safetyEvaluation aur safety Also calledIse yeh bhi kehte hain: Interpretability, XAI 🔊

🧸 Think of it likeAise samjho: Like a doctor who doesn't just say 'take rest', but explains 'your fever and these test results are why'.Jaise doctor sirf 'aaram karo' nahi kehta, balki samjhaata hai 'aapka bukhaar aur yeh test results iski wajah hain'.

In lending, healthcare and hiring, people and regulators need reasons. Tools like SHAP show which features pushed a prediction up or down; simpler models are easier to explain. For LLMs, showing sources (citations) is a practical form of explainability.Lending, healthcare aur hiring mein logon aur regulators ko wajah chahiye. SHAP jaise tools dikhate hain kaunse features ne prediction upar ya neeche kiya; simple models samjhaane mein aasaan hain. LLMs ke liye sources (citations) dikhana explainability ka practical roop hai.

💡 Real exampleAsli example: 'Loan declined mainly because of 3 missed payments and high EMI to income ratio.''Loan mainly 3 miss hue payments aur zyaada EMI to income ratio ki wajah se reject hua.'

📖 Learn itSeekho: Feature engineering: turning raw data into useful inputsFeature engineering: raw data ko kaam ke inputs mein badalna · Enterprise AI: integration, security and complianceEnterprise AI: integration, security aur compliance

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Bias (in data and models) · Feature

Machine learningMachine learning 🔊

🧸 Think of it likeAise samjho: Like judging a fisherman on two things at once: how many of the fish in the lake he caught, and how much of his catch was actually fish and not old shoes.Machhuaare ko do cheezon pe ek saath parakhna: jheel ki kitni machhliyaan pakdi, aur jo pakda usmein kitna sach mein machhli tha, purane joote nahi.

F1 is the harmonic mean of precision and recall, so it's high only when both are reasonably high. It's handy for comparing models on imbalanced data, but in real decisions you usually pick a threshold based on which mistake costs more, rather than maximising F1 blindly.F1 precision aur recall ka harmonic mean hai, isliye yeh tabhi ooncha hota hai jab dono theek theek oonche hon. Imbalanced data pe models compare karne mein kaam ka hai, par asli faislon mein aksar threshold is hisaab se chunte ho ki kaunsi galti mehngi hai, F1 ko aankh band karke badhaane ke bajaye.

💡 Real exampleAsli example: Precision 0.8 and recall 0.6 give F1 of about 0.69.Precision 0.8 aur recall 0.6 se F1 lagbhag 0.69.

📖 Learn itSeekho: Measuring models: accuracy, precision and recallModels ko measure karna: accuracy, precision aur recall

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Precision · Recall

Machine learningMachine learning Also calledIse yeh bhi kehte hain: Input variable 🔊

🧸 Think of it likeAise samjho: Like the clues a property dealer uses to guess a flat's price: size, floor, area and distance from the station.Jaise property dealer flat ki keemat andaaza lagaane ke liye clues use karta hai: size, floor, area aur station se doori.

Features are the columns a model learns from; the label is the answer it predicts. Good features carry real signal and are available at prediction time. Raw data usually needs to be turned into features: ratios, counts over time windows, encoded categories.Features woh columns hain jinse model seekhta hai; label woh jawaab hai jo woh predict karta hai. Acche features mein asli signal hota hai aur prediction ke time available hote hain. Raw data ko aksar features mein badalna padta hai: ratios, time windows mein ginti, encoded categories.

💡 Real exampleAsli example: For house prices: area, bedrooms, distance to metro station.Ghar ki keemat ke liye: area, bedrooms, metro station se doori.

📖 Learn itSeekho: How machines learn: features, labels and trainingMachines kaise seekhti hain: features, labels aur training · Feature engineering: turning raw data into useful inputsFeature engineering: raw data ko kaam ke inputs mein badalna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Feature engineering · Label · Data leakage

Machine learningMachine learning 🔊

🧸 Think of it likeAise samjho: Like a cook turning raw ingredients into ready masalas: the model gets inputs that are much more useful than the raw data.Jaise cook kachche saamaan ko ready masale mein badalta hai: model ko raw data se kahin zyada kaam ke inputs milte hain.

Examples: EMI to income ratio, days since last purchase, number of failed logins in the last hour, one hot encoded city. Good feature engineering often beats a fancier model on tabular data. Compute features the same way in training and serving, and check them for leakage.Examples: EMI to income ratio, last purchase se kitne din, pichhle ghante mein fail hue logins, one hot encoded city. Tabular data pe accha feature engineering aksar fancy model ko harata hai. Training aur serving mein features ek hi tarike se banao, aur leakage check karo.

💡 Real exampleAsli example: Creating 'average order value in the last 30 days' from raw order rows.Raw order rows se 'pichhle 30 din ka average order value' banana.

📖 Learn itSeekho: Feature engineering: turning raw data into useful inputsFeature engineering: raw data ko kaam ke inputs mein badalna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Feature · Feature store · Training serving skew

MLOps and deploymentMLOps aur deployment 🔊

🧸 Think of it likeAise samjho: Like a shared pantry of ready made masalas, so every cook uses exactly the same mix in training and in the real kitchen.Ready made masalon ki shared pantry jaisa, taaki har cook training aur asli kitchen dono mein bilkul same mix use kare.

Teams compute features once and reuse them, with point in time correctness for training (no peeking into the future) and low latency lookup for serving. This prevents training serving skew. Examples include Feast and the feature stores in cloud ML platforms.Teams features ek baar banati hain aur dobara use karti hain, training ke liye point in time sahi (future mein jhaankna nahi) aur serving ke liye low latency lookup ke saath. Isse training serving skew rukta hai. Examples: Feast aur cloud ML platforms ke feature stores.

💡 Real exampleAsli example: Fraud scoring reads 'transactions in the last 10 minutes' from the feature store in milliseconds.Fraud scoring feature store se milliseconds mein 'pichhle 10 minute ki transactions' padhta hai.

📖 Learn itSeekho: Training vs inference: how ML runs in productionTraining vs inference: ML production mein kaise chalta hai · Streaming data: Kafka and real-time pipelinesStreaming data: Kafka aur real-time pipelines

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Feature · Training serving skew

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: In context learning 🔊

🧸 Think of it likeAise samjho: Like showing a new helper three folded shirts before asking them to fold the rest.Naye helper ko teen tah ki hui shirts dikha kar baaki tah karne ko kehna.

Instead of training, you include 2 to 5 input and output examples, then the new input. The model picks up the format and style. It's a fast way to improve consistency for classification, extraction and tone. Zero shot means giving instructions with no examples.Training ke bajaye 2 se 5 input output examples daalo, phir naya input. Model format aur style pakad leta hai. Classification, extraction aur tone mein consistency badhaane ka tez tarika hai. Zero shot matlab bina examples sirf instructions.

💡 Real exampleAsli example: Showing three tickets with their categories, then asking for the category of a fourth.Teen tickets unki categories ke saath dikha kar chauthe ki category poochhna.

📖 Learn itSeekho: Prompting: giving clear instructions to an LLMPrompting: LLM ko saaf instructions dena

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Prompt engineering · Zero shot

LLMs and GenAILLMs aur GenAI 🔊

🧸 Think of it likeAise samjho: Like a general doctor doing a short extra course to become a skin specialist: the basics stay, the focus sharpens.Jaise general doctor chhota extra course karke skin specialist banta hai: basics wahi rehte hain, focus tez hota hai.

Fine tuning teaches a style, format or narrow skill, for example always answering in a fixed JSON format or writing in a brand's tone. It needs hundreds to thousands of high quality examples and an eval set. It's not the best way to add changing facts; use RAG for knowledge. LoRA makes fine tuning cheap.Fine tuning style, format ya ek narrow skill sikhaata hai, jaise hamesha fixed JSON format mein jawaab ya brand ke tone mein likhna. Iske liye sainkdon se hazaaron high quality examples aur eval set chahiye. Badalte facts jodne ka yeh best tarika nahi; knowledge ke liye RAG use karo. LoRA fine tuning ko sasta banata hai.

💡 Real exampleAsli example: Fine tuning a small model to turn doctor notes into a standard discharge format.Doctor notes ko standard discharge format mein badalne ke liye chhota model fine tune karna.

📖 Learn itSeekho: Fine-tuning and LoRA: specialising a modelFine-tuning aur LoRA: model ko specialise karna · AI literacy for decisions: build vs buy, cost and limitsDecisions ke liye AI literacy: build vs buy, cost aur limits

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: LoRA · RAG (retrieval augmented generation) · Instruction tuning

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: Base model, frontier model 🔊

🧸 Think of it likeAise samjho: Like a graduate with broad general knowledge who can be trained quickly for many different jobs.Ek graduate jaisa jiske paas broad general knowledge hai aur jise jaldi se kai alag jobs ke liye train kiya ja sakta hai.

Foundation models (LLMs, vision and multimodal models) are trained once at huge cost, then used through prompting, RAG or fine tuning for specific products. 'Frontier' models are the most capable at a given time. Most companies build on top of foundation models rather than training their own.Foundation models (LLMs, vision aur multimodal models) ek baar bahut kharch se train hote hain, phir prompting, RAG ya fine tuning se specific products ke liye use hote hain. 'Frontier' models kisi samay ke sabse capable models hote hain. Zyaadatar companies apna train karne ke bajaye foundation models ke upar banati hain.

💡 Real exampleAsli example: Claude, GPT and Gemini, and open models like Llama, are foundation models.Claude, GPT aur Gemini, aur Llama jaise open models, foundation models hain.

📖 Learn itSeekho: What is an LLM? Tokens, context and temperatureLLM kya hai? Tokens, context aur temperature · AI literacy for decisions: build vs buy, cost and limitsDecisions ke liye AI literacy: build vs buy, cost aur limits

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: LLM (large language model) · Fine tuning · Open weights model

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: Generative AI 🔊

🧸 Think of it likeAise samjho: Like a helper who doesn't just find an existing answer but writes, draws or composes something new for you.Ek helper jaisa jo sirf maujooda jawaab nahi dhoondhta, balki aapke liye kuch naya likhta, banata ya compose karta hai.

Generative models learn the patterns of their training data well enough to produce new examples. LLMs generate text and code; diffusion models generate images. GenAI products include writing assistants, copilots, chatbots and design tools, and they need evaluation and guardrails because outputs vary.Generative models apne training data ke patterns itne acche se seekhte hain ki naye examples bana sakein. LLMs text aur code banate hain; diffusion models images. GenAI products mein writing assistants, copilots, chatbots aur design tools hain, aur unhe evaluation aur guardrails chahiye kyunki outputs badalte rehte hain.

💡 Real exampleAsli example: Drafting a product description from a few bullet points.Kuch bullet points se product description draft karna.

📖 Learn itSeekho: What is AI? AI vs ML vs Deep Learning vs GenAIAI kya hai? AI vs ML vs Deep Learning vs GenAI · What is an LLM? Tokens, context and temperatureLLM kya hai? Tokens, context aur temperature

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: LLM (large language model) · Diffusion model · Hallucination

MLOps and deploymentMLOps aur deployment Also calledIse yeh bhi kehte hain: Graphics processing unit 🔊

🧸 Think of it likeAise samjho: Like 1,000 students each solving one small sum at the same time, instead of one genius solving 1,000 sums one by one.Jaise 1,000 students ek saath ek ek chhota sawaal solve karein, ek genius ke 1,000 sawaal ek ek karke solve karne ki jagah.

Neural networks are mostly matrix multiplications, which GPUs do far faster than CPUs. GPU memory limits how big a model and batch can be. GPUs are expensive, so serving techniques like batching, quantisation and caching matter for cost.Neural networks zyaadatar matrix multiplications hain, jo GPUs CPUs se kahin tez karte hain. GPU memory tay karti hai model aur batch kitne bade ho sakte hain. GPUs mehenge hain, isliye batching, quantisation aur caching jaise serving tarike cost ke liye zaroori hain.

💡 Real exampleAsli example: Fine tuning a 7B model with QLoRA on one 24 GB GPU.Ek 24 GB GPU pe QLoRA se 7B model fine tune karna.

📖 Learn itSeekho: Serving LLMs efficiently: GPUs, batching and costLLMs ko efficiently serve karna: GPUs, batching aur cost

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Quantisation · Batch size · Inference

Deep learningDeep learning 🔊

🧸 Think of it likeAise samjho: Like feeling which way the ground slopes under your feet: it tells you which direction makes the error go up, so you step the other way.Pairon ke neeche zameen ka dhalaan mehsoos karne jaisa: batata hai kis taraf error badhta hai, toh aap ulti taraf kadam rakhte ho.

For each weight, the gradient says how much the loss would change if that weight changed slightly. Training uses gradients to update all weights at once. Backpropagation computes them efficiently.Har weight ke liye gradient batata hai ki weight thoda badalne pe loss kitna badlega. Training gradients se saare weights ek saath update karti hai. Backpropagation inhe efficiently nikaalta hai.

💡 Real exampleAsli example: If increasing a weight raises the error, its gradient is positive, so training decreases it.Agar weight badhaane se error badhe, to uska gradient positive hai, isliye training use ghataati hai.

📖 Learn itSeekho: Gradients and gradient descent: how models improveGradients aur gradient descent: models behtar kaise hote hain

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Gradient descent · Backpropagation · Loss function

Machine learningMachine learning Also calledIse yeh bhi kehte hain: XGBoost, LightGBM, CatBoost 🔊

🧸 Think of it likeAise samjho: Like a student who, after each mock test, studies only the questions they got wrong. Many small rounds add up to a strong result.Jaise student har mock test ke baad sirf galat sawaalon ko padhta hai. Kai chhote rounds milkar strong result dete hain.

Gradient boosting is usually the strongest choice for tabular business data: fraud, churn, credit risk, pricing. Key settings are the number of trees, depth and learning rate; use early stopping on validation data. Libraries: XGBoost, LightGBM, CatBoost.Tabular business data ke liye gradient boosting aksar sabse mazboot choice hai: fraud, churn, credit risk, pricing. Zaroori settings: trees ki ginti, depth aur learning rate; validation data pe early stopping lagao. Libraries: XGBoost, LightGBM, CatBoost.

💡 Real exampleAsli example: An XGBoost model on 40 features raising fraud recall from 0.62 to 0.81.40 features pe XGBoost model se fraud recall 0.62 se 0.81.

📖 Learn itSeekho: Decision trees, random forests and boostingDecision trees, random forests aur boosting

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Decision tree · Random forest · Early stopping

Deep learningDeep learning Also calledIse yeh bhi kehte hain: SGD, Adam 🔊

🧸 Think of it likeAise samjho: Like walking downhill in fog: you can't see the bottom, so you keep taking small steps where the ground slopes down.Kohre mein pahaad se neeche utarne jaisa: neeche dikhta nahi, toh jahan dhalaan neeche ho wahan chhote kadam lete raho.

Compute the loss, compute gradients, move each weight a little against its gradient, repeat. The learning rate sets the step size. Stochastic gradient descent uses mini batches; optimisers like Adam adapt the step for each weight and are common defaults.Loss nikalo, gradients nikalo, har weight ko uske gradient ke ulta thoda khiskaao, repeat. Learning rate kadam ka size tay karta hai. Stochastic gradient descent mini batches use karta hai; Adam jaise optimisers har weight ka kadam adjust karte hain aur common defaults hain.

💡 Real exampleAsli example: Like walking downhill in fog by always stepping where the ground slopes down.Jaise kohre mein hamesha dhalaan ki taraf kadam rakh kar pahaad se utarna.

📖 Learn itSeekho: Gradients and gradient descent: how models improveGradients aur gradient descent: models behtar kaise hote hain

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Gradient · Learning rate · Loss function

LLMs and GenAILLMs aur GenAI 🔊

🧸 Think of it likeAise samjho: Like an open book exam: answer from the book in front of you, not from fuzzy memory.Open book exam jaisa: saamne rakhi kitaab se jawaab do, dhundhli yaad se nahi.

Grounded answers come from retrieved documents, databases or APIs placed in the prompt, ideally with citations. Grounding reduces hallucinations and keeps answers current. Instruct the model to answer only from the sources and to say when they don't contain the answer.Grounded jawaab prompt mein daale gaye retrieved documents, databases ya APIs se aate hain, behtar ho to citations ke saath. Grounding hallucinations kam karti hai aur jawaab current rakhti hai. Model ko kaho sirf sources se jawaab de aur bataaye jab sources mein jawaab na ho.

💡 Real exampleAsli example: Answering 'what is my leave balance?' by reading the HR system, not guessing.'Meri leave balance kitni hai?' ka jawaab andaaze se nahi, HR system padh kar dena.

📖 Learn itSeekho: RAG: letting an LLM answer from your documentsRAG: LLM ko aapke documents se jawaab dene dena · AI safety in products: hallucinations, prompt injection and guardrailsProducts mein AI safety: hallucinations, prompt injection aur guardrails

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: RAG (retrieval augmented generation) · Hallucination · Citations

Product and careerProduct aur career 🔊

🧸 Think of it likeAise samjho: Like dieting to lose weight while making sure your energy levels don't crash: one number improves, another must not get worse.Jaise weight kam karne ke liye diet karte hue dhyaan rakhna ki energy na gire: ek number sudhre, doosra kharaab na ho.

In an A/B test or launch, the main metric might be conversion or time saved, while guardrails watch complaints, errors, latency, cost and fairness. A 'win' that breaks a guardrail is not a win.A/B test ya launch mein main metric conversion ya bacha time ho sakta hai, jabki guardrails complaints, errors, latency, cost aur fairness dekhte hain. Jo 'jeet' guardrail tode woh jeet nahi.

💡 Real exampleAsli example: AI replies cut handling time by 20%, but customer complaints must not rise.AI replies handling time 20% kam karte hain, par customer complaints nahi badhni chahiye.

📖 Learn itSeekho: Metrics and A/B testing for AI featuresAI features ke liye metrics aur A/B testing · Writing an AI product spec (PRD) with an eval planEval plan ke saath AI product spec (PRD) likhna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: A/B test · Metric

Evaluation and safetyEvaluation aur safety 🔊

🧸 Think of it likeAise samjho: Like the railings on a mountain road: you can still drive freely, but they stop you going over the edge.Pahaadi sadak ki railing jaisa: aap aaraam se chala sakte ho, par woh kinaare se girne nahi deti.

Guardrails include input filters (prompt injection, abuse), output checks (personal data, policy violations, valid format), tool limits (permissions, max refund amount), agent limits (max steps, budgets) and human approval for risky actions. Important rules should be enforced in code, not only in the prompt.Guardrails mein input filters (prompt injection, abuse), output checks (personal data, policy ka ullanghan, valid format), tool limits (permissions, maximum refund amount), agent limits (max steps, budgets) aur risky actions pe human approval aate hain. Zaroori rules code mein lagne chahiye, sirf prompt mein nahi.

💡 Real exampleAsli example: A refund agent can't issue more than ₹2,000 without a person approving.Refund agent bina insaan ke approval ₹2,000 se zyaada nahi de sakta.

📖 Learn itSeekho: AI safety in products: hallucinations, prompt injection and guardrailsProducts mein AI safety: hallucinations, prompt injection aur guardrails · AI agents: plan, act, observe, repeatAI agents: plan, act, observe, repeat

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Prompt injection · Human in the loop · Alignment

Evaluation and safetyEvaluation aur safety Also calledIse yeh bhi kehte hain: Confabulation 🔊

🧸 Think of it likeAise samjho: Like a student who doesn't know the answer but writes a confident, made up one to fill the page.Jaise student ko jawaab nahi aata par page bharne ke liye confident hokar mann se likh deta hai.

LLMs generate plausible text rather than looking up facts, so they can invent citations, numbers or policies. Reduce it with grounding (RAG), instructions to say 'I don't know', citations, lower temperature for factual tasks, code checks, and human review for high stakes. Measure the hallucination rate with evals.LLMs facts dekhne ke bajaye sahi lagne wala text banate hain, isliye citations, numbers ya policies gadh sakte hain. Grounding (RAG), 'mujhe nahi pata' kehne ke instructions, citations, factual kaam mein kam temperature, code checks, aur high stakes mein human review se kam karo. Evals se hallucination rate naapo.

💡 Real exampleAsli example: A chatbot inventing a refund policy that the company never had.Chatbot ka aisi refund policy gadh dena jo company ki kabhi thi hi nahi.

📖 Learn itSeekho: AI safety in products: hallucinations, prompt injection and guardrailsProducts mein AI safety: hallucinations, prompt injection aur guardrails · What is an LLM? Tokens, context and temperatureLLM kya hai? Tokens, context aur temperature

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Grounding · RAG (retrieval augmented generation) · Evals (evaluation)

Evaluation and safetyEvaluation aur safety Also calledIse yeh bhi kehte hain: HITL 🔊

🧸 Think of it likeAise samjho: Like a junior writing a letter and the manager reading and signing it before it is sent.Jaise junior letter likhta hai aur bhejne se pehle manager padh kar sign karta hai.

For risky or regulated decisions, AI drafts or suggests and a human approves, edits or rejects. Human feedback also becomes training and evaluation data. As quality is proven, the share of cases needing review can be reduced.Risky ya regulated faislon mein AI draft ya suggest karta hai aur insaan approve, edit ya reject karta hai. Insaani feedback training aur evaluation data bhi banta hai. Quality sabit hone pe review wale cases ka hissa kam kiya ja sakta hai.

💡 Real exampleAsli example: A doctor signs every AI drafted discharge summary.Doctor har AI drafted discharge summary sign karta hai.

📖 Learn itSeekho: UX for AI features: designing for uncertaintyAI features ke liye UX: uncertainty ke hisaab se design · AI safety in products: hallucinations, prompt injection and guardrailsProducts mein AI safety: hallucinations, prompt injection aur guardrails

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Guardrails · Agent (AI agent)

LLMs and GenAILLMs aur GenAI 🔊

🧸 Think of it likeAise samjho: Like asking both a librarian who knows exact book titles and one who understands what you mean, then combining their picks.Jaise do librarian se poochna, ek jo exact book titles jaanta hai aur ek jo aapka matlab samajhta hai, phir dono ki pasand milaana.

Vector search understands meaning but misses exact codes and names; keyword search (BM25) catches exact terms but misses paraphrases. Hybrid search runs both and merges the rankings, and a reranker often refines the top results. It's a common upgrade for RAG.Vector search matlab samajhta hai par exact codes aur naam miss karta hai; keyword search (BM25) exact terms pakadta hai par alag shabdon mein kahi baat miss karta hai. Hybrid search dono chalata hai aur rankings milata hai, aur reranker aksar top results ko aur behtar karta hai. RAG ke liye common upgrade hai.

💡 Real exampleAsli example: Finding both 'leave policy for new parents' and the exact form code 'HR-112'.'Naye parents ki leave policy' aur exact form code 'HR-112' dono dhoondhna.

📖 Learn itSeekho: RAG: letting an LLM answer from your documentsRAG: LLM ko aapke documents se jawaab dene dena · Embeddings and vector databases: searching by meaningEmbeddings aur vector databases: matlab se search karna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: BM25 (keyword search) · Semantic search · Reranking

Machine learningMachine learning 🔊

🧸 Think of it likeAise samjho: Like the oven temperature and timer you set before baking: the cake doesn't choose them, you do.Bake karne se pehle set kiya oven ka temperature aur timer: cake inhe nahi chunta, aap chunte ho.

Weights are learned; hyperparameters are chosen by you and tuned on validation data, by hand, grid search, random search or tools like Optuna. Never tune on the test set.Weights seekhe jaate hain; hyperparameters aap chunte ho aur validation data pe tune karte ho, haath se, grid search, random search ya Optuna jaise tools se. Test set pe kabhi tune mat karo.

💡 Real exampleAsli example: Trying max depth 4, 6 and 8 for a gradient boosting model.Gradient boosting model ke liye max depth 4, 6 aur 8 try karna.

📖 Learn itSeekho: Train/test split and overfittingTrain/test split aur overfitting

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Validation set · Parameter

DataData Also calledIse yeh bhi kehte hain: Idempotent 🔊

🧸 Think of it likeAise samjho: Like pressing a lift button twice: the lift still comes once.Lift ka button do baar dabaane jaisa: lift phir bhi ek hi baar aati hai.

Pipelines are retried and backfilled all the time. Writing by overwriting a date partition or upserting on a unique key, instead of blindly appending, makes reruns safe and prevents double counting. In APIs, idempotency keys stop duplicate payments when a request is retried.Pipelines baar baar retry aur backfill hoti hain. Bina soche append ke bajaye date partition overwrite ya unique key pe upsert karke likhne se reruns safe hote hain aur double counting rukti hai. APIs mein idempotency keys request retry hone pe duplicate payments rokti hain.

💡 Real exampleAsli example: Rerunning yesterday's sales load replaces yesterday's rows instead of adding them again.Kal ka sales load dobara chalaane pe kal ki rows badal jaati hain, dobara judti nahi.

📖 Learn itSeekho: Data pipelines: ETL, Airflow, dbt and SparkData pipelines: ETL, Airflow, dbt aur Spark

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Data pipeline · ETL and ELT

MLOps and deploymentMLOps aur deployment Also calledIse yeh bhi kehte hain: Prediction, serving 🔊

🧸 Think of it likeAise samjho: Training is studying; inference is sitting the exam: using what was learned to answer a new question.Training padhai hai; inference exam dena hai: jo seekha use naye sawaal ka jawaab dene mein use karna.

Training happens occasionally; inference happens every time a user needs a prediction or an LLM answer. Inference must be fast, reliable and affordable, so it involves APIs, autoscaling, batching, caching and monitoring. For most companies inference is the main ongoing AI cost.Training kabhi kabhi hoti hai; inference har baar hota hai jab user ko prediction ya LLM jawaab chahiye. Inference tez, reliable aur sasta hona chahiye, isliye isme APIs, autoscaling, batching, caching aur monitoring aate hain. Zyaadatar companies ke liye inference hi AI ka main chalta kharch hai.

💡 Real exampleAsli example: Each time you ask a chatbot a question, the model runs inference.Har baar chatbot se sawaal poochhne pe model inference chalata hai.

📖 Learn itSeekho: Training vs inference: how ML runs in productionTraining vs inference: ML production mein kaise chalta hai · Serving LLMs efficiently: GPUs, batching and costLLMs ko efficiently serve karna: GPUs, batching aur cost

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Training · Latency · Batch inference

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: Supervised fine tuning, SFT 🔊

🧸 Think of it likeAise samjho: Like teaching a well read person good manners at a help desk: listen to the request, then answer it clearly.Jaise ek padhe likhe insaan ko help desk ke tareeke sikhaana: request suno, phir saaf jawaab do.

A base model only continues text. Instruction tuning teaches it to respond to questions and tasks in a helpful format, and preference methods like RLHF or DPO refine it further. Chat models you use are instruction tuned.Base model sirf text aage badhaata hai. Instruction tuning use sawaalon aur kaamon ka madadgaar format mein jawaab dena sikhaati hai, aur RLHF ya DPO jaise preference methods use aur sudhaarte hain. Aap jo chat models use karte ho woh instruction tuned hain.

💡 Real exampleAsli example: Training on pairs like 'Summarise this email' → a good summary.'Summarise this email' → accha summary jaise jodon pe training.

📖 Learn itSeekho: Fine-tuning and LoRA: specialising a modelFine-tuning aur LoRA: model ko specialise karna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Fine tuning · RLHF · Alignment

Machine learningMachine learning 🔊

🧸 Think of it likeAise samjho: Like placing 4 ice cream carts in a park so each visitor walks to the nearest one, then moving the carts until each sits in the middle of its crowd.Jaise park mein 4 ice cream carts rakhna taaki har visitor sabse paas wali tak jaaye, phir carts ko tab tak khiskaana jab tak har ek apni bheed ke beech na aa jaaye.

Pick K, place K centres, assign every item to the nearest centre, move each centre to the average of its items, and repeat until stable. It's simple and fast but you must choose K and it prefers round, similar sized clusters.K chuno, K centres rakho, har item ko sabse paas ke centre ko do, har centre ko uske items ke average pe le jao, aur stable hone tak repeat karo. Simple aur tez hai par K chunna padta hai aur yeh gol, ek jaise size ke clusters pasand karta hai.

💡 Real exampleAsli example: Grouping 10,000 customers into 4 spending segments.10,000 customers ko 4 spending segments mein baantna.

📖 Learn itSeekho: How machines learn: features, labels and trainingMachines kaise seekhti hain: features, labels aur training

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Clustering · Unsupervised learning

MLOps and deploymentMLOps aur deployment Also calledIse yeh bhi kehte hain: K8s 🔊

🧸 Think of it likeAise samjho: Like a building manager for containers: decides which flat each one lives in, adds more when it gets busy and replaces any that break.Containers ke liye building manager jaisa: tay karta hai kaun kis flat mein rahe, bheed hone pe aur jodta hai aur toote hue badal deta hai.

You describe what you want (for example 4 replicas of the model API with 2 GB memory each) and Kubernetes keeps it that way: it restarts failed containers, spreads traffic through Services, rolls out new versions gradually and autoscales. It's widely used for model serving and GPU workloads, though managed services are simpler for small projects.Aap batate ho kya chahiye (jaise model API ki 4 replicas, har ek 2 GB memory ke saath) aur Kubernetes waisa hi banaye rakhta hai: fail hue containers restart karta, Services se traffic baantta, naye versions dheere rollout karta aur autoscale karta hai. Model serving aur GPU workloads ke liye kaafi use hota hai, par chhote projects ke liye managed services simple hain.

💡 Real exampleAsli example: A crashed model pod is replaced automatically within seconds.Crash hua model pod seconds mein apne aap badal diya jaata hai.

📖 Learn itSeekho: Kubernetes: running containers at scaleKubernetes: containers ko bade scale pe chalaana

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Docker · Autoscaling · Deployment

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: Key value cache 🔊

🧸 Think of it likeAise samjho: Like keeping notes while reading a long story, so for each new page you don't have to reread the whole book.Lambi kahaani padhte waqt notes banaane jaisa, taaki har naye page ke liye poori kitaab dobara na padhni pade.

Generating text one token at a time would otherwise redo work for all earlier tokens at each step. The KV cache makes generation much faster but uses GPU memory that grows with context length and batch size, which inference engines manage with techniques like paged attention.Ek ek token generate karne mein warna har step pe pichhle saare tokens ka kaam dobara hota. KV cache generation kaafi tez karta hai par GPU memory leta hai jo context length aur batch size ke saath badhti hai, jise inference engines paged attention jaise tarikon se sambhaalte hain.

💡 Real exampleAsli example: Long chats use more GPU memory per user because the KV cache grows.Lambi chats mein har user ki zyaada GPU memory lagti hai kyunki KV cache badhta hai.

📖 Learn itSeekho: Transformers and attention: how LLMs readTransformers aur attention: LLMs kaise padhte hain · Serving LLMs efficiently: GPUs, batching and costLLMs ko efficiently serve karna: GPUs, batching aur cost

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Attention (self attention) · Inference · GPU

DataData Also calledIse yeh bhi kehte hain: Target, ground truth 🔊

🧸 Think of it likeAise samjho: The answer written on the back of a flashcard.Flashcard ke peeche likha jawaab.

In a churn dataset the label is 'cancelled: yes or no'; in price prediction it's the price. Labels can be noisy or delayed (fraud labels arrive weeks later via chargebacks). Label quality limits model quality.Churn dataset mein label 'cancel kiya: haan ya nahi' hai; price prediction mein price. Labels mein noise ya deri ho sakti hai (fraud labels hafton baad chargebacks se aate hain). Label ki quality model ki quality ki seema tay karti hai.

💡 Real exampleAsli example: Each past email marked 'spam' or 'not spam'.Har purana email 'spam' ya 'not spam' mark kiya hua.

📖 Learn itSeekho: How machines learn: features, labels and trainingMachines kaise seekhti hain: features, labels aur training

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Feature · Annotation (labelling) · Supervised learning

MLOps and deploymentMLOps aur deployment Also calledIse yeh bhi kehte hain: Response time, p95 🔊

🧸 Think of it likeAise samjho: Like the time between ordering a chai and getting it in your hand.Chai order karne aur haath mein aane ke beech ka time.

Measure percentiles, not just averages: p95 latency is the time within which 95% of requests finish. For LLMs, track time to first token (how soon text appears) and tokens per second. Reduce latency with streaming, shorter prompts, caching, smaller models and batching.Sirf average nahi, percentiles naapo: p95 latency woh time hai jisme 95% requests poori hoti hain. LLMs ke liye time to first token (text kitni jaldi dikhta hai) aur tokens per second track karo. Streaming, chhote prompts, caching, chhote models aur batching se latency kam karo.

💡 Real exampleAsli example: p95 latency of 800 ms means 95 of 100 requests finish within 0.8 seconds.p95 latency 800 ms matlab 100 mein se 95 requests 0.8 second mein poori hoti hain.

📖 Learn itSeekho: Serving LLMs efficiently: GPUs, batching and costLLMs ko efficiently serve karna: GPUs, batching aur cost · Serving a model as an API with FastAPIModel ko FastAPI se API ke roop mein serve karna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Streaming (responses) · Inference · Autoscaling

Deep learningDeep learning 🔊

🧸 Think of it likeAise samjho: Like the size of your steps when looking for something in the dark: too big and you keep overshooting, too small and it takes forever.Andhere mein kuch dhoondhte waqt kadmon ke size jaisa: bahut bade toh aage nikal jaate ho, bahut chhote toh bahut der lagti hai.

Too high and training jumps around or diverges; too low and it's painfully slow or gets stuck. Schedules (warm up, then decay) and optimisers like Adam help. It's usually the most important hyperparameter to tune.Bahut zyaada to training uchhalti ya bigad jaati hai; bahut kam to bahut slow ya atak jaati hai. Schedules (warm up, phir decay) aur Adam jaise optimisers madad karte hain. Aam taur pe tune karne ka sabse zaroori hyperparameter yahi hai.

💡 Real exampleAsli example: Loss bouncing up and down usually means the learning rate is too high.Loss ka upar neeche uchhalna aksar matlab learning rate bahut zyaada hai.

📖 Learn itSeekho: Gradients and gradient descent: how models improveGradients aur gradient descent: models behtar kaise hote hain

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Gradient descent · Hyperparameter

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: Language model 🔊

🧸 Think of it likeAise samjho: Like a super powered autocomplete that has read a huge library, so it can continue any text in a sensible way: answers, emails, code.Ek super powerful autocomplete jaisa jisne ek bahut badi library padh li hai, isliye kisi bhi text ko samajhdaari se aage likh sakta hai: jawaab, emails, code.

LLMs are transformers with billions of parameters, trained to predict the next token and then tuned to follow instructions. They power chatbots, copilots, agents and text processing. They work within a context window, cost money per token, and can hallucinate, so products add RAG, tools, evals and guardrails.LLMs arabon parameters wale transformers hain, agla token predict karne ke liye train hue aur phir instructions maanne ke liye tune hue. Inse chatbots, copilots, agents aur text processing chalte hain. Yeh context window mein kaam karte hain, har token ka paisa lagta hai, aur hallucinate kar sakte hain, isliye products RAG, tools, evals aur guardrails jodte hain.

💡 Real exampleAsli example: Claude, GPT, Gemini and Llama are LLMs.Claude, GPT, Gemini aur Llama LLMs hain.

📖 Learn itSeekho: What is an LLM? Tokens, context and temperatureLLM kya hai? Tokens, context aur temperature · Transformers and attention: how LLMs readTransformers aur attention: LLMs kaise padhte hain

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Token · Transformer · Context window · Hallucination

Evaluation and safetyEvaluation aur safety Also calledIse yeh bhi kehte hain: Model graded eval 🔊

🧸 Think of it likeAise samjho: Like a senior teacher using a marking scheme to grade answer sheets written by another teacher's class.Jaise senior teacher marking scheme se doosri class ki answer sheets check kare.

It scales evaluation of open ended answers (helpfulness, faithfulness, tone). Judges have biases (prefer longer answers, the first option, their own style), so give a clear rubric, ask for reasoning, check agreement with human ratings, and use code checks when the answer can be verified exactly.Isse open ended jawaabon (helpfulness, faithfulness, tone) ka evaluation bade scale pe hota hai. Judges ke biases hain (lambe jawaab, pehla option, apna style pasand), isliye saaf rubric do, reasoning maango, human ratings se agreement check karo, aur jab jawaab exactly verify ho sake tab code checks use karo.

💡 Real exampleAsli example: A judge model scoring each RAG answer 1 to 5 for 'supported by the sources'.Judge model har RAG jawaab ko 'sources se supported' ke liye 1 se 5 score deta hai.

📖 Learn itSeekho: Evaluating AI: how to know it actually worksAI ko evaluate karna: kaise pata chale ki sach mein kaam karta hai

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Evals (evaluation) · Benchmark

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: Low rank adaptation, QLoRA 🔊

🧸 Think of it likeAise samjho: Like adding a small clip on lens to your phone camera instead of buying a whole new camera.Naya camera kharidne ki jagah phone camera pe ek chhota clip on lens lagaane jaisa.

The base weights are frozen; LoRA learns a small low rank update for some layers, often under 1% of parameters. This cuts GPU memory and storage, and you can swap adapters per task. QLoRA keeps the base model in 4 bit precision so large models can be tuned on a single GPU.Base weights freeze rehte hain; LoRA kuch layers ke liye chhota low rank update seekhta hai, aksar 1% se kam parameters. Isse GPU memory aur storage kam lagti hai, aur har task ke liye adapters badal sakte ho. QLoRA base model ko 4 bit precision mein rakhta hai taaki bade models ek GPU pe tune ho sakein.

💡 Real exampleAsli example: A 30 MB LoRA adapter that makes a 7B model write in your company's report format.30 MB ka LoRA adapter jo 7B model ko company ke report format mein likhwata hai.

📖 Learn itSeekho: Fine-tuning and LoRA: specialising a modelFine-tuning aur LoRA: model ko specialise karna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Fine tuning · Quantisation · Parameter

Deep learningDeep learning Also calledIse yeh bhi kehte hain: Cost function, objective 🔊

🧸 Think of it likeAise samjho: Like the 'how far off was I?' score after each guess. Training keeps adjusting to make that score smaller.Har guess ke baad 'kitna galat tha?' wala score. Training use chhota karne ke liye adjust karti rehti hai.

Regression often uses mean squared error; classification uses cross entropy. The loss must be differentiable so gradients can be computed. The loss you optimise isn't always the business metric you care about, so evaluate with both.Regression aksar mean squared error use karta hai; classification cross entropy. Loss differentiable hona chahiye taaki gradients nikal sakein. Jo loss optimise karte ho woh hamesha woh business metric nahi jiski parwah hai, isliye dono se evaluate karo.

💡 Real exampleAsli example: Predicting ₹48 lakh for a ₹50 lakh flat gives a small squared error loss.₹50 lakh ke flat ke liye ₹48 lakh predict karna chhota squared error loss deta hai.

📖 Learn itSeekho: Gradients and gradient descent: how models improveGradients aur gradient descent: models behtar kaise hote hain

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Cross entropy · Gradient descent

LLMs and GenAILLMs aur GenAI 🔊

🧸 Think of it likeAise samjho: Like remembering the first and last songs of a long playlist but not the ones in the middle.Lambi playlist ke pehle aur aakhri gaane yaad rehna par beech wale nahi.

When you stuff many retrieved chunks into a long context, relevant facts in the middle can be overlooked. Mitigations: retrieve fewer, better chunks, rerank so the most relevant come first, and keep prompts focused.Jab bahut saare retrieved chunks lambe context mein thoos dete ho, beech ke relevant facts nazarandaaz ho sakte hain. Upaay: kam par behtar chunks laao, rerank karo taaki sabse relevant pehle aayein, aur prompts focused rakho.

💡 Real exampleAsli example: The right policy paragraph is chunk 14 of 25, and the model answers from chunk 2 instead.Sahi policy paragraph 25 mein se chunk 14 hai, aur model chunk 2 se jawaab de deta hai.

📖 Learn itSeekho: RAG: letting an LLM answer from your documentsRAG: LLM ko aapke documents se jawaab dene dena

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Context window · Reranking · Chunking

AI basicsAI basics Also calledIse yeh bhi kehte hain: ML 🔊

🧸 Think of it likeAise samjho: Like learning to spot ripe mangoes by looking at hundreds of them, instead of someone writing you a rule book.Jaise sau mango dekh kar pakka aam pehchaanna seekhna, kisi ke rule book likh kar dene ki jagah.

You give a learning algorithm data (features and, for supervised learning, labels); it finds a function that predicts well on new data. Main types are supervised (predict a known target), unsupervised (find structure) and reinforcement learning (learn from rewards). Deep learning is machine learning with large neural networks.Learning algorithm ko data dete ho (features aur, supervised learning mein, labels); woh aisa function dhoondhta hai jo naye data pe accha predict kare. Main types: supervised (known target predict karna), unsupervised (structure dhoondhna) aur reinforcement learning (rewards se seekhna). Deep learning bade neural networks ke saath machine learning hai.

💡 Real exampleAsli example: Learning to predict house prices from thousands of past sales.Hazaaron purani bikri se ghar ki keemat predict karna seekhna.

📖 Learn itSeekho: What is AI? AI vs ML vs Deep Learning vs GenAIAI kya hai? AI vs ML vs Deep Learning vs GenAI · How machines learn: features, labels and trainingMachines kaise seekhti hain: features, labels aur training

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: AI (artificial intelligence) · Deep learning · Supervised learning

Agents and toolsAgents aur tools 🔊

🧸 Think of it likeAise samjho: Like a universal charger port: any app that supports it can plug into any tool that supports it.Universal charger port jaisa: jo bhi app ise support kare, woh kisi bhi support karne wale tool mein plug ho sakta hai.

A tool owner builds one MCP server that exposes tools, resources and prompts; any MCP compatible client (assistants, IDEs, agents) can use it. This avoids writing custom integrations for every app. Servers still need authentication, permissions and careful review, because they can read data and take actions.Tool ka owner ek MCP server banata hai jo tools, resources aur prompts deta hai; koi bhi MCP compatible client (assistants, IDEs, agents) use use kar sakta hai. Isse har app ke liye custom integration likhna bachta hai. Servers ko phir bhi authentication, permissions aur dhyaan se review chahiye, kyunki woh data padh aur actions le sakte hain.

💡 Real exampleAsli example: One MCP server for your ticketing system lets several AI assistants create and search tickets.Ticketing system ka ek MCP server kai AI assistants ko tickets banane aur search karne deta hai.

📖 Learn itSeekho: MCP: a universal plug for AI toolsMCP: AI tools ke liye universal plug

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Tool calling · Agent (AI agent)

Product and careerProduct aur career Also calledIse yeh bhi kehte hain: KPI 🔊

🧸 Think of it likeAise samjho: Like the speedometer and fuel gauge in a car: numbers you watch to know how things are going.Gaadi ke speedometer aur fuel gauge jaisa: numbers jinhe dekh kar pata chalta hai sab kaisa chal raha hai.

For AI features use layers: model quality (accuracy, faithfulness), user outcomes (time saved, task success), adoption (usage, retention) and business results (revenue, cost per ticket), plus guardrails. Pick metrics before building and avoid relying on one number.AI features ke liye layers use karo: model quality (accuracy, faithfulness), user outcomes (bacha time, task success), adoption (usage, retention) aur business results (revenue, per ticket cost), saath mein guardrails. Banane se pehle metrics chuno aur ek number pe nirbhar mat raho.

💡 Real exampleAsli example: Edit distance between AI draft and final text, a measure of draft quality.AI draft aur final text ke beech edit distance, draft quality ka naap.

📖 Learn itSeekho: Metrics and A/B testing for AI featuresAI features ke liye metrics aur A/B testing · Writing an AI product spec (PRD) with an eval planEval plan ke saath AI product spec (PRD) likhna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Guardrail metric · Evals (evaluation) · A/B test

MLOps and deploymentMLOps aur deployment Also calledIse yeh bhi kehte hain: LLMOps 🔊

🧸 Think of it likeAise samjho: Like running a restaurant kitchen, not just inventing a recipe: steady supplies, quality checks and fixing things fast when they go wrong.Sirf recipe banana nahi, restaurant ki kitchen chalaane jaisa: lagaataar saamaan, quality check aur kuch bigde toh jaldi theek karna.

MLOps covers reproducible training, experiment tracking, model registries, CI/CD for models, deployment, monitoring for drift and performance, and retraining. LLMOps adds prompt versioning, eval suites, cost tracking, tracing and safety controls.MLOps mein reproducible training, experiment tracking, model registries, models ke liye CI/CD, deployment, drift aur performance ki monitoring, aur retraining aate hain. LLMOps prompt versioning, eval suites, cost tracking, tracing aur safety controls jodta hai.

💡 Real exampleAsli example: Automatically retraining a demand model monthly and promoting it only if it beats the current one.Demand model ko har mahine apne aap retrain karna aur sirf tab promote karna jab current ko hara de.

📖 Learn itSeekho: MLOps: experiment tracking and model registriesMLOps: experiment tracking aur model registries · Monitoring models: drift and silent failuresModels ki monitoring: drift aur chupke se hone wale failures

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Model registry · Monitoring · CI/CD

AI basicsAI basics 🔊

🧸 Think of it likeAise samjho: Like a recipe the computer worked out by itself from examples: put inputs in, get a prediction out.Ek recipe jaisa jo computer ne examples se khud banayi: inputs daalo, prediction niklo.

In ML, a model is an algorithm plus its learned parameters (weights), saved to a file and loaded to make predictions. 'Model' can mean a small logistic regression or a huge LLM. In production it travels with its preprocessing and a version number.ML mein model ek algorithm plus uske seekhe hue parameters (weights) hai, jo file mein save hota hai aur predictions ke liye load hota hai. 'Model' chhota logistic regression bhi ho sakta hai aur bahut bada LLM bhi. Production mein yeh apne preprocessing aur version number ke saath chalta hai.

💡 Real exampleAsli example: fraud_model_v3.pkl, a trained XGBoost model with its feature pipeline.fraud_model_v3.pkl, apni feature pipeline ke saath trained XGBoost model.

📖 Learn itSeekho: How machines learn: features, labels and trainingMachines kaise seekhti hain: features, labels aur training

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Parameter · Training · Inference

MLOps and deploymentMLOps aur deployment 🔊

🧸 Think of it likeAise samjho: Like a library catalogue for models: which version exists, how good it was, and which one is on the shelf right now.Models ke liye library catalogue jaisa: kaunsa version hai, kitna achha tha, aur abhi shelf pe kaunsa hai.

Each registered model links to the training run, data version and evaluation results. Promotion to production is explicit, so you always know what's live and can roll back in one step. MLflow and cloud ML platforms provide registries.Har registered model training run, data version aur evaluation results se juda hota hai. Production mein promotion saaf hota hai, isliye hamesha pata hota hai kya live hai aur ek kadam mein rollback ho sakta hai. MLflow aur cloud ML platforms registries dete hain.

💡 Real exampleAsli example: churn-model v12 in production, v13 in staging awaiting approval.churn-model v12 production mein, v13 staging mein approval ke intezaar mein.

📖 Learn itSeekho: MLOps: experiment tracking and model registriesMLOps: experiment tracking aur model registries

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: MLOps · CI/CD

MLOps and deploymentMLOps aur deployment Also calledIse yeh bhi kehte hain: Observability 🔊

🧸 Think of it likeAise samjho: Like the health monitor beside a hospital bed that beeps when something goes wrong.Hospital bed ke paas wale health monitor jaisa jo kuch galat hone pe beep karta hai.

For models: system metrics (latency, errors), data (missing values, drift), predictions (distribution, confidence), real performance when labels arrive, and business metrics. For LLM apps: traces of every step, token cost, feedback and eval scores. Each alert should map to an action.Models ke liye: system metrics (latency, errors), data (missing values, drift), predictions (distribution, confidence), labels aane pe asli performance, aur business metrics. LLM apps ke liye: har step ke traces, token cost, feedback aur eval scores. Har alert ka ek action hona chahiye.

💡 Real exampleAsli example: An alert fires when 20% of a feature's values are suddenly missing.Ek feature ki 20% values achanak missing hone pe alert bajta hai.

📖 Learn itSeekho: Monitoring models: drift and silent failuresModels ki monitoring: drift aur chupke se hone wale failures

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Data drift · Concept drift · Latency

LLMs and GenAILLMs aur GenAI 🔊

🧸 Think of it likeAise samjho: Like a person who can read, look at pictures and listen at the same time, and connect them all.Ek insaan jaisa jo ek saath padh sakta hai, tasveer dekh sakta hai aur sun sakta hai, aur sabko jod sakta hai.

Multimodal LLMs can read a screenshot, describe a photo, transcribe speech or answer questions about a chart. This enables apps like invoice reading, visual inspection and voice assistants.Multimodal LLMs screenshot padh sakte hain, photo ka varnan, awaaz likh sakte hain ya chart ke baare mein sawaalon ke jawaab de sakte hain. Isse invoice padhna, visual inspection aur voice assistants jaisi apps banti hain.

💡 Real exampleAsli example: Uploading a photo of a bill and asking 'what's the total GST?'Bill ki photo upload karke poochhna 'total GST kitna hai?'

📖 Learn itSeekho: What is an LLM? Tokens, context and temperatureLLM kya hai? Tokens, context aur temperature

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Computer vision · LLM (large language model)

Deep learningDeep learning 🔊

🧸 Think of it likeAise samjho: Like a big team passing notes: each member adds a little judgement, and together they reach an answer. Practice tunes how much they trust each other's notes.Ek badi team jaisa jo notes aage badhaati hai: har member thoda faisla jodta hai, aur milkar jawaab tak pahunchte hain. Practice se tay hota hai kaun kiske notes pe kitna bharosa kare.

Each neuron takes a weighted sum of inputs and applies an activation function; layers stack these to learn complex patterns. Networks are trained with backpropagation and gradient descent. Variants include convolutional networks for images and transformers for language.Har neuron inputs ka weighted sum leta hai aur activation function lagata hai; layers inhe jod kar complex patterns seekhti hain. Networks backpropagation aur gradient descent se train hote hain. Variants: images ke liye convolutional networks aur bhasha ke liye transformers.

💡 Real exampleAsli example: A network that takes pixel values and outputs 'cat' or 'dog'.Network jo pixel values leta hai aur 'cat' ya 'dog' deta hai.

📖 Learn itSeekho: Neural networks: layers that learnNeural networks: layers jo seekhti hain

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Deep learning · Activation function · Backpropagation

AI basicsAI basics 🔊

🧸 Think of it likeAise samjho: Teaching computers to read, understand and write the way people talk.Computers ko log jaise baat karte hain waise padhna, samajhna aur likhna sikhaana.

Classic NLP used rules and statistical models for tasks like sentiment analysis and named entity recognition. Today most NLP is done with transformers and LLMs, including Indian languages and Hinglish.Classic NLP sentiment analysis aur named entity recognition jaise kaamon ke liye rules aur statistical models use karta tha. Aaj zyaadatar NLP transformers aur LLMs se hota hai, Indian languages aur Hinglish samet.

💡 Real exampleAsli example: Detecting whether a product review is positive or negative.Pata lagaana ki product review positive hai ya negative.

📖 Learn itSeekho: What is an LLM? Tokens, context and temperatureLLM kya hai? Tokens, context aur temperature

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: LLM (large language model) · Token

Deep learningDeep learning Also calledIse yeh bhi kehte hain: Scaling, LayerNorm 🔊

🧸 Think of it likeAise samjho: Like converting marks out of 50 and out of 100 to percentages before comparing them.50 mein se aur 100 mein se marks ko compare karne se pehle percentage mein badalna.

For classic ML, scale features (for example to mean 0 and standard deviation 1) so large valued features don't dominate; fit the scaler on training data only. Inside neural networks, layers like BatchNorm and LayerNorm keep activations stable; transformers use LayerNorm.Classic ML mein features scale karo (jaise mean 0 aur standard deviation 1) taaki badi values wale features haavi na hon; scaler sirf training data pe fit karo. Neural networks ke andar BatchNorm aur LayerNorm jaisi layers activations stable rakhti hain; transformers LayerNorm use karte hain.

💡 Real exampleAsli example: Scaling salary (₹ lakhs) and age (years) so both matter fairly in a distance based model.Salary (₹ lakh) aur umar (saal) ko scale karna taaki distance based model mein dono barabar matter karein.

📖 Learn itSeekho: Feature engineering: turning raw data into useful inputsFeature engineering: raw data ko kaam ke inputs mein badalna · Neural networks: layers that learnNeural networks: layers jo seekhti hain

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Feature engineering · Neural network

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: Open source model, Llama, Mistral 🔊

🧸 Think of it likeAise samjho: Like a recipe that's been published: anyone can cook it in their own kitchen and adjust the spices.Chhapi hui recipe jaisa: koi bhi apni kitchen mein bana sakta hai aur masale badal sakta hai.

Open weights models can run on your own servers, which helps with data privacy, cost at high volume and customisation. You take on hosting, scaling and safety work yourself. Check the licence: some allow commercial use with conditions.Open weights models aapke apne servers pe chal sakte hain, jo data privacy, zyaada volume pe cost aur customisation mein madad karta hai. Hosting, scaling aur safety ka kaam aapko khud karna padta hai. Licence check karo: kuch commercial use shart ke saath allow karte hain.

💡 Real exampleAsli example: Running a Llama model inside a bank's own data centre so customer data never leaves.Bank ke apne data centre mein Llama model chalana taaki customer data bahar na jaaye.

📖 Learn itSeekho: AI literacy for decisions: build vs buy, cost and limitsDecisions ke liye AI literacy: build vs buy, cost aur limits · Serving LLMs efficiently: GPUs, batching and costLLMs ko efficiently serve karna: GPUs, batching aur cost

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Foundation model · Fine tuning · Quantisation

Machine learningMachine learning 🔊

🧸 Think of it likeAise samjho: Like a student who mugs up last year's question paper: perfect on those questions, lost on new ones.Jaise student pichhle saal ka paper rat leta hai: unhi sawaalon pe perfect, naye sawaalon pe kho jaata hai.

Signs: very high training score, much lower validation score. Causes: a model too complex for the data, too little data, training too long. Fixes: more or cleaner data, simpler models, regularisation, dropout, early stopping and cross validation.Lakshan: training score bahut ooncha, validation score kaafi kam. Wajah: data ke hisaab se bahut complex model, kam data, bahut der training. Upaay: zyaada ya saaf data, simple models, regularisation, dropout, early stopping aur cross validation.

💡 Real exampleAsli example: 98% training accuracy but 65% on the test set.98% training accuracy par test set pe 65%.

📖 Learn itSeekho: Train/test split and overfittingTrain/test split aur overfitting

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Underfitting · Regularisation · Bias variance trade off

Deep learningDeep learning Also calledIse yeh bhi kehte hain: Weights 🔊

🧸 Think of it likeAise samjho: Like the tiny knobs on a giant music mixer: training turns billions of them until the output sounds right.Ek bade music mixer ke chhote knobs jaisa: training arabon knobs ghumaati hai jab tak awaaz sahi na aaye.

Parameters (weights and biases) are what the model 'knows'. More parameters can capture more patterns but need more data, memory and compute. '7B' means about 7 billion parameters. Hyperparameters, by contrast, are settings you choose.Parameters (weights aur biases) hi model ki 'jaankari' hain. Zyaada parameters zyaada patterns pakad sakte hain par zyaada data, memory aur compute chahiye. '7B' matlab lagbhag 7 arab parameters. Hyperparameters, iske ulat, aapki chuni settings hain.

💡 Real exampleAsli example: A 7B parameter model needs about 14 GB of memory in 16 bit precision.7B parameter model ko 16 bit precision mein lagbhag 14 GB memory chahiye.

📖 Learn itSeekho: Neural networks: layers that learnNeural networks: layers jo seekhti hain · Serving LLMs efficiently: GPUs, batching and costLLMs ko efficiently serve karna: GPUs, batching aur cost

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Hyperparameter · Quantisation · Model

Machine learningMachine learning 🔊

🧸 Think of it likeAise samjho: When the smoke alarm rings, how often is there really a fire? High precision means few false alarms.Jab smoke alarm bajta hai, kitni baar sach mein aag hoti hai? High precision matlab kam jhoothe alarm.

Precision = true positives ÷ (true positives + false positives). High precision means few false alarms, which matters when false alarms are costly, like blocking genuine payments or sending real emails to spam. Raising the threshold usually increases precision and lowers recall.Precision = true positives ÷ (true positives + false positives). High precision matlab kam false alarms, jo tab zaroori hai jab false alarm mehenga ho, jaise asli payments block karna ya asli emails spam mein daalna. Threshold badhaane se aam taur pe precision badhti aur recall ghatta hai.

💡 Real exampleAsli example: The model flags 20 payments; 8 are real fraud, so precision is 40%.Model 20 payments flag karta hai; 8 asli fraud hain, to precision 40%.

📖 Learn itSeekho: Measuring models: accuracy, precision and recallModels ko measure karna: accuracy, precision aur recall

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Recall · F1 score · Confusion matrix · Threshold

LLMs and GenAILLMs aur GenAI 🔊

🧸 Think of it likeAise samjho: Like the instructions you give a new helper: what to do, the background they need, and how you want it done.Naye helper ko diye instructions jaisa: kya karna hai, zaroori background, aur kaise chahiye.

A prompt usually has a system part (role, rules, format) and user messages, plus any retrieved documents or tool results. Clear goals, needed context, explicit constraints, examples and an output format produce much better results. Treat prompts like code: version and test them.Prompt mein aam taur pe system hissa (role, rules, format) aur user messages hote hain, saath mein retrieved documents ya tool results. Saaf goals, zaroori context, saaf constraints, examples aur output format se kahin behtar results milte hain. Prompts ko code ki tarah treat karo: version aur test karo.

💡 Real exampleAsli example: 'Summarise this complaint in 2 lines and suggest a reply in a polite tone.''Is complaint ka 2 line mein summary do aur vinamra tone mein reply suggest karo.'

📖 Learn itSeekho: Prompting: giving clear instructions to an LLMPrompting: LLM ko saaf instructions dena

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Prompt engineering · System prompt · Few shot prompting

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: Prompt design 🔊

🧸 Think of it likeAise samjho: Like learning how to brief a helper so well that they get it right the first time, every time.Helper ko itni achhi tarah samjhaana seekhna ki woh har baar pehli baar mein sahi kare.

Techniques: clear role and goal, relevant context, step by step instructions, few shot examples, output schemas, asking for citations, and telling the model what to do when unsure. Measure changes with an eval set rather than a few manual tries.Tarike: saaf role aur goal, relevant context, step by step instructions, few shot examples, output schemas, citations maangna, aur model ko batana ki unsure ho to kya kare. Kuch haath ki koshishon ke bajaye eval set se badlaav naapo.

💡 Real exampleAsli example: Adding 'If the answer isn't in the sources, say you don't know' cut wrong answers by half.'Agar jawaab sources mein nahi hai, to bolo pata nahi' jodne se galat jawaab aadhe ho gaye.

📖 Learn itSeekho: Prompting: giving clear instructions to an LLMPrompting: LLM ko saaf instructions dena · Evaluating AI: how to know it actually worksAI ko evaluate karna: kaise pata chale ki sach mein kaam karta hai

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Prompt · Evals (evaluation) · Chain of thought

Evaluation and safetyEvaluation aur safety Also calledIse yeh bhi kehte hain: Jailbreak 🔊

🧸 Think of it likeAise samjho: Like a note slipped into a stack of letters saying 'ignore your boss and send me the money'. A careful assistant must not obey it.Jaise chitthiyon ke dher mein ek parchi chhupi ho 'boss ki baat mat suno, mujhe paise bhejo'. Samajhdaar assistant ko yeh nahi maanna chahiye.

Direct injection comes from the user ('ignore your rules'); indirect injection hides instructions in web pages, emails or files the AI reads. There's no complete fix, so defend in layers: treat external text as data, limit tool permissions, require confirmation for risky actions, filter outputs and test with known attacks.Direct injection user se aata hai ('apne rules bhool jao'); indirect injection un web pages, emails ya files mein chhupe instructions hain jo AI padhta hai. Poora fix nahi hai, isliye layers mein bachaav: bahar ke text ko data maano, tool permissions seemit rakho, risky actions pe confirmation, outputs filter karo aur known attacks se test karo.

💡 Real exampleAsli example: A resume with hidden white text saying 'rate this candidate 10/10'.Resume mein chhupa safed text 'is candidate ko 10/10 do'.

📖 Learn itSeekho: AI safety in products: hallucinations, prompt injection and guardrailsProducts mein AI safety: hallucinations, prompt injection aur guardrails

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Guardrails · Agent (AI agent)

Deep learningDeep learning 🔊

🧸 Think of it likeAise samjho: Like a well stocked workshop with ready tools for building and training neural networks in Python.Ek bhari hui workshop jaisa jismein Python mein neural networks banane aur train karne ke ready tools hain.

PyTorch provides tensors (arrays on GPUs), automatic gradients, layers and optimisers. Most research and many production models use it, and the Hugging Face ecosystem builds on it. TensorFlow and JAX are alternatives.PyTorch tensors (GPUs pe arrays), automatic gradients, layers aur optimisers deta hai. Zyaadatar research aur kai production models ise use karte hain, aur Hugging Face ecosystem isi pe bana hai. TensorFlow aur JAX alternatives hain.

💡 Real exampleAsli example: Writing a training loop for a text classifier in 30 lines of PyTorch.PyTorch ki 30 lines mein text classifier ka training loop likhna.

📖 Learn itSeekho: Neural networks: layers that learnNeural networks: layers jo seekhti hain

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Deep learning · GPU

MLOps and deploymentMLOps aur deployment Also calledIse yeh bhi kehte hain: 4 bit, 8 bit, INT8 🔊

🧸 Think of it likeAise samjho: Like saving a photo in a smaller file size: it looks almost the same but takes far less space.Photo ko chhote file size mein save karne jaisa: dikhti lagbhag same hai par jagah bahut kam leti hai.

Instead of 16 bit numbers, weights are stored in 8 or 4 bits, cutting memory by 2 to 4 times with usually small quality loss. It lets bigger models fit on cheaper GPUs or even laptops. Always re-check quality on your eval set after quantising.16 bit numbers ke bajaye weights 8 ya 4 bits mein rakhe jaate hain, memory 2 se 4 guna kam, aam taur pe thode quality loss ke saath. Isse bade models saste GPUs ya laptops pe bhi fit hote hain. Quantise karne ke baad hamesha apne eval set pe quality dobara check karo.

💡 Real exampleAsli example: A 14 GB model becomes about 4 GB in 4 bit form.14 GB ka model 4 bit roop mein lagbhag 4 GB ka ho jaata hai.

📖 Learn itSeekho: Serving LLMs efficiently: GPUs, batching and costLLMs ko efficiently serve karna: GPUs, batching aur cost · Fine-tuning and LoRA: specialising a modelFine-tuning aur LoRA: model ko specialise karna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: GPU · LoRA · Parameter

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: Retrieval augmented generation 🔊

🧸 Think of it likeAise samjho: Like an open book exam: first find the right pages in the book, then write the answer from them and mention the page.Open book exam jaisa: pehle kitaab mein sahi pages dhoondho, phir unse jawaab likho aur page number batao.

RAG has two phases. Indexing: split documents into chunks, create embeddings and store them in a vector database with metadata and permissions. Answering: embed the question, retrieve the best chunks (often hybrid search plus reranking), put them in the prompt with instructions to answer only from them and cite sources, then generate. RAG keeps answers current and grounded without retraining. Most failures are retrieval failures, so measure retrieval and answer quality separately.RAG ke do phases hain. Indexing: documents ko chunks mein todo, embeddings banao aur metadata aur permissions ke saath vector database mein rakho. Answering: sawaal ka embedding banao, best chunks nikaalo (aksar hybrid search plus reranking), unhe prompt mein daalo is instruction ke saath ki sirf unhi se jawaab de aur sources cite kare, phir generate karo. RAG bina retrain jawaab current aur grounded rakhta hai. Zyaadatar failures retrieval ke hote hain, isliye retrieval aur jawaab ki quality alag naapo.

💡 Real exampleAsli example: An HR bot that answers 'how many casual leaves do I get?' from the latest policy PDF, with the page reference.HR bot jo 'mujhe kitni casual leaves milti hain?' ka jawaab latest policy PDF se, page reference ke saath deta hai.

📖 Learn itSeekho: RAG: letting an LLM answer from your documentsRAG: LLM ko aapke documents se jawaab dene dena · Embeddings and vector databases: searching by meaningEmbeddings aur vector databases: matlab se search karna · Evaluating AI: how to know it actually worksAI ko evaluate karna: kaise pata chale ki sach mein kaam karta hai

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Embedding · Vector database · Chunking · Hybrid search · Grounding

Machine learningMachine learning 🔊

🧸 Think of it likeAise samjho: Like asking 500 people for directions and going with the majority: one person may be wrong, the crowd usually isn't.500 logon se raasta poochh kar majority ki baat maanna: ek insaan galat ho sakta hai, bheed aksar nahi.

Each tree sees a random subset of rows and features, so trees make different mistakes, and averaging cancels many of them out. Random forests are robust, need little tuning and are a strong baseline for tabular data; gradient boosting is often a bit more accurate.Har tree rows aur features ka random hissa dekhta hai, isliye trees alag galtiyaan karte hain, aur average karne se kai galtiyaan cancel ho jaati hain. Random forests mazboot hain, kam tuning chahiye aur tabular data ke liye accha baseline hain; gradient boosting aksar thoda zyaada accurate hota hai.

💡 Real exampleAsli example: 500 trees voting on whether a customer will churn.500 trees ka vote ki customer churn karega ya nahi.

📖 Learn itSeekho: Decision trees, random forests and boostingDecision trees, random forests aur boosting

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Decision tree · Gradient boosting

Machine learningMachine learning Also calledIse yeh bhi kehte hain: Sensitivity, true positive rate 🔊

🧸 Think of it likeAise samjho: Of all the real fires, how many did the alarm catch? High recall means few fires are missed.Saari asli aag mein se alarm ne kitni pakdi? High recall matlab kam aag chhooti.

Recall = true positives ÷ (true positives + false negatives). High recall means few misses, which matters when missing a case is costly, like cancer screening or fraud. Lowering the threshold usually increases recall and lowers precision.Recall = true positives ÷ (true positives + false negatives). High recall matlab kam miss, jo tab zaroori hai jab case miss karna mehenga ho, jaise cancer screening ya fraud. Threshold ghataane se aam taur pe recall badhta aur precision ghatti hai.

💡 Real exampleAsli example: There were 10 frauds and the model caught 8: recall is 80%.10 fraud the aur model ne 8 pakde: recall 80%.

📖 Learn itSeekho: Measuring models: accuracy, precision and recallModels ko measure karna: accuracy, precision aur recall

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Precision · F1 score · Threshold

Machine learningMachine learning 🔊

🧸 Think of it likeAise samjho: Like guessing the price of a second hand car from its age, brand and kilometres: the answer is a number.Purani gaadi ki keemat uski umar, brand aur kilometre se andaaza lagaane jaisa: jawaab ek number hai.

Regression models output continuous values. Start with linear regression as a baseline, then try gradient boosting or neural networks. Evaluate with errors in real units: MAE (average absolute error), RMSE (punishes big errors) or MAPE (percentage error).Regression models continuous values dete hain. Baseline ke roop mein linear regression se shuru karo, phir gradient boosting ya neural networks try karo. Asli units mein errors se evaluate karo: MAE (average absolute error), RMSE (bade errors ko zyaada saza) ya MAPE (percent error).

💡 Real exampleAsli example: Predicting a flat's price as ₹82 lakh when it sold for ₹80 lakh: an error of ₹2 lakh.Flat ki keemat ₹82 lakh predict karna jab woh ₹80 lakh mein bika: ₹2 lakh ka error.

📖 Learn itSeekho: Two kinds of prediction: regression and classificationDo tarah ki prediction: regression aur classification

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Classification · Loss function

Machine learningMachine learning Also calledIse yeh bhi kehte hain: L1, L2, weight decay 🔊

🧸 Think of it likeAise samjho: Like a word limit on an essay: it forces you to keep only what matters instead of stuffing in every detail.Essay ki word limit jaisa: aapko sirf zaroori cheezein rakhne pe majboor karta hai, har detail thoonsne ki jagah.

Common forms: penalising large weights (L2 or ridge, weight decay), pushing some weights to zero (L1 or lasso), dropout in neural networks, limiting tree depth, and early stopping. Its strength is a hyperparameter tuned on validation data.Common roop: bade weights pe penalty (L2 ya ridge, weight decay), kuch weights ko zero tak dhakelna (L1 ya lasso), neural networks mein dropout, tree depth seemit karna, aur early stopping. Iski taakat ek hyperparameter hai jo validation data pe tune hota hai.

💡 Real exampleAsli example: Adding weight decay to stop a model fitting random noise in a small dataset.Chhote dataset mein random noise fit hone se rokne ke liye weight decay jodna.

📖 Learn itSeekho: Train/test split and overfittingTrain/test split aur overfitting

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Overfitting · Dropout · Early stopping

Machine learningMachine learning Also calledIse yeh bhi kehte hain: RL 🔊

🧸 Think of it likeAise samjho: Like training a dog with treats: good actions get a reward, so the dog learns to repeat them.Kutte ko treats se train karne jaisa: achhe kaam pe inaam milta hai, toh woh unhe dohraana seekhta hai.

The agent observes a state, takes an action, gets a reward, and gradually learns a policy that maximises total reward. It powers game playing systems and robotics, and variants like RLHF are used to make LLMs more helpful.Agent ek state dekhta hai, action leta hai, reward paata hai, aur dheere dheere aisi policy seekhta hai jo kul reward zyaada kare. Yeh game khelne wale systems aur robotics chalaata hai, aur RLHF jaise variants LLMs ko zyaada madadgaar banane mein use hote hain.

💡 Real exampleAsli example: A program learning to play chess by playing millions of games against itself.Program jo khud se laakhon games khel kar chess seekhta hai.

📖 Learn itSeekho: What is AI? AI vs ML vs Deep Learning vs GenAIAI kya hai? AI vs ML vs Deep Learning vs GenAI

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: RLHF · Supervised learning

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: Reranker, cross encoder 🔊

🧸 Think of it likeAise samjho: Like shortlisting 50 job applicants quickly, then interviewing them carefully to pick the best 5.Jaise 50 applicants ko jaldi shortlist karna, phir dhyaan se interview karke best 5 chunna.

First retrieval quickly finds, say, 50 candidates; a reranker model reads the question and each candidate together and orders them by true relevance, keeping the best few for the LLM. It adds some latency but usually improves RAG answers a lot.Pehla retrieval jaldi, maan lo, 50 candidates dhoondhta hai; reranker model sawaal aur har candidate ko saath padh kar asli relevance se order karta hai, aur LLM ke liye kuch best rakhta hai. Thodi latency judti hai par aksar RAG jawaab kaafi behtar hote hain.

💡 Real exampleAsli example: Reordering 50 search hits so the exact policy clause moves from position 12 to position 1.50 search hits ko aise reorder karna ki exact policy clause position 12 se 1 pe aa jaaye.

📖 Learn itSeekho: RAG: letting an LLM answer from your documentsRAG: LLM ko aapke documents se jawaab dene dena

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Hybrid search · RAG (retrieval augmented generation) · Lost in the middle

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: Reinforcement learning from human feedback 🔊

🧸 Think of it likeAise samjho: Like a trainee chef cooking two versions, a panel of tasters picking the better one, and the chef slowly learning what people like.Jaise trainee chef do version banaye, tasters ka panel behtar wala chune, aur chef dheere dheere seekhe ki logon ko kya pasand hai.

People compare pairs of model answers; a reward model learns these preferences; the LLM is then optimised with reinforcement learning to produce higher rewarded answers. It improves helpfulness and safety. DPO is a simpler alternative that skips the separate reward model.Log model ke jawaabon ke jodon ko compare karte hain; reward model yeh pasand seekhta hai; phir LLM ko reinforcement learning se zyaada reward wale jawaab dene ke liye optimise kiya jaata hai. Isse helpfulness aur safety sudharti hai. DPO simple alternative hai jo alag reward model chhod deta hai.

💡 Real exampleAsli example: Raters prefer a clear, polite answer over a rambling one; the model learns that style.Raters lambi baaton ke bajaye saaf, vinamra jawaab pasand karte hain; model woh style seekhta hai.

📖 Learn itSeekho: Fine-tuning and LoRA: specialising a modelFine-tuning aur LoRA: model ko specialise karna · AI safety in products: hallucinations, prompt injection and guardrailsProducts mein AI safety: hallucinations, prompt injection aur guardrails

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: DPO · Alignment · Reinforcement learning

DataData Also calledIse yeh bhi kehte hain: Selection bias 🔊

🧸 Think of it likeAise samjho: Like judging a whole city's favourite food by only asking people outside one pizza shop.Poore shehar ka pasandeeda khaana sirf ek pizza dukaan ke bahar logon se poochh kar tay karna.

If you only survey current users, you miss people who left; if you train only on approved loans, you never see how rejected applicants would have behaved. Biased samples lead to wrong conclusions and unfair models. Think about who is missing from your data.Agar sirf current users ka survey karo, to chhod kar gaye log chhoot jaate hain; agar sirf approved loans pe train karo, to kabhi nahi dekhte ki reject hue log kaise behave karte. Biased samples galat nateeje aur anyaayi models dete hain. Socho aapke data se kaun gayab hai.

💡 Real exampleAsli example: An app review study only reading 5 star reviews.App reviews ki study jo sirf 5 star reviews padhti hai.

📖 Learn itSeekho: Correlation vs causation, and sampling biasCorrelation vs causation, aur sampling bias

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Bias (in data and models) · Dataset

Deep learningDeep learning 🔊

🧸 Think of it likeAise samjho: Like knowing that more practice hours and better coaching improve a runner's time in a predictable way.Jaise pata hona ki zyada practice ghante aur behtar coaching runner ka time ek tay andaaz mein sudhaarte hain.

Research found that loss falls smoothly as a power law with more parameters, training tokens and compute, and that model size and data should grow together for the best use of compute. These laws guide how big labs plan training runs.Research mein paaya gaya ki zyaada parameters, training tokens aur compute ke saath loss power law ki tarah smoothly girta hai, aur compute ka best use karne ke liye model size aur data saath badhne chahiye. Yeh laws bade labs ko training runs plan karne mein guide karte hain.

💡 Real exampleAsli example: Doubling training data and model size together gives a predictable drop in loss.Training data aur model size dono double karne pe loss mein anumaanit girawat.

📖 Learn itSeekho: Reading and reproducing research papersResearch papers padhna aur reproduce karna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Parameter · Foundation model

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: Vector search 🔊

🧸 Think of it likeAise samjho: Like a shopkeeper who understands 'something for a headache' and hands you the right medicine, even though you never said its name.Jaise dukaandaar 'sar dard ke liye kuch' samajh kar sahi dawai de deta hai, bhale aapne naam nahi bola.

Documents and queries are turned into embeddings, and the closest vectors are returned. It finds relevant results even when wording differs. Combine it with keyword search for exact identifiers.Documents aur queries embeddings mein badle jaate hain, aur sabse paas ke vectors lautaaye jaate hain. Shabd alag hon tab bhi relevant results milte hain. Exact identifiers ke liye ise keyword search ke saath milao.

💡 Real exampleAsli example: Searching 'my card is not working abroad' finds 'international transactions disabled'.'my card is not working abroad' search karne pe 'international transactions disabled' milta hai.

📖 Learn itSeekho: Embeddings and vector databases: searching by meaningEmbeddings aur vector databases: matlab se search karna · RAG: letting an LLM answer from your documentsRAG: LLM ko aapke documents se jawaab dene dena

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Embedding · Hybrid search · Vector database

Deep learningDeep learning 🔊

🧸 Think of it likeAise samjho: Like turning cricket match scores into win chances that add up to 100%.Cricket match ke scores ko jeet ke chances mein badalna jo milkar 100% hote hain.

Softmax exponentiates each score and divides by the total, so bigger scores get most of the probability. It's used in classification outputs, in attention weights, and when an LLM picks the next token; temperature reshapes it.Softmax har score ka exponential leta hai aur total se divide karta hai, isliye bade scores ko zyaadatar probability milti hai. Yeh classification outputs, attention weights, aur LLM ke agla token chunne mein use hota hai; temperature ise reshape karta hai.

💡 Real exampleAsli example: Scores 2, 1, 0 become probabilities of about 0.67, 0.24 and 0.09.Scores 2, 1, 0 lagbhag 0.67, 0.24 aur 0.09 probabilities ban jaate hain.

📖 Learn itSeekho: Neural networks: layers that learnNeural networks: layers jo seekhti hain · Transformers and attention: how LLMs readTransformers aur attention: LLMs kaise padhte hain

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Cross entropy · Temperature · Attention (self attention)

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: Token streaming 🔊

🧸 Think of it likeAise samjho: Like watching someone type a message live instead of waiting for the full message to arrive.Jaise kisi ko live message type karte dekhna, poora message aane ka intezaar karne ki jagah.

The first words of an answer are ready much sooner than the whole answer. Streaming sends chunks to the browser (server sent events or WebSockets) so the product feels fast. Run validations after the stream ends and support a stop button. (Data streaming, like Kafka, is a different idea: processing events continuously.)Jawaab ke pehle shabd poore jawaab se kaafi pehle taiyaar hote hain. Streaming chunks browser ko bhejta hai (server sent events ya WebSockets) taaki product tez lage. Validations stream khatam hone ke baad chalao aur stop button do. (Data streaming, jaise Kafka, alag idea hai: events ko lagaataar process karna.)

💡 Real exampleAsli example: ChatGPT style typing effect while the answer is generated.Jawaab bante waqt ChatGPT jaisa typing effect.

📖 Learn itSeekho: Calling an LLM API: keys, requests and costsLLM API call karna: keys, requests aur costs · Streaming data: Kafka and real-time pipelinesStreaming data: Kafka aur real-time pipelines

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Latency · LLM (large language model)

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: JSON mode, schema 🔊

🧸 Think of it likeAise samjho: Like asking someone to fill a form with fixed boxes instead of writing a free letter, so a computer can read it.Kisi se khuli chitthi ki jagah fixed boxes wala form bharwaana, taaki computer padh sake.

Use the provider's structured output or JSON mode with a schema, keep fields simple, set low temperature, and always validate in code (for example with Pydantic). Retry once with the error message, then fall back safely.Provider ka structured output ya schema wala JSON mode use karo, fields simple rakho, temperature kam rakho, aur hamesha code mein validate karo (jaise Pydantic se). Error message ke saath ek baar retry, phir safe fallback.

💡 Real exampleAsli example: Extracting {invoice_no, date, total, gst} from a scanned invoice.Scanned invoice se {invoice_no, date, total, gst} nikaalna.

📖 Learn itSeekho: Calling an LLM API: keys, requests and costsLLM API call karna: keys, requests aur costs · Tool use (function calling): letting an LLM take actionsTool use (function calling): LLM ko action lene dena

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Tool calling · Prompt engineering

Machine learningMachine learning 🔊

🧸 Think of it likeAise samjho: Like learning with a teacher who marks every answer right or wrong.Aise teacher ke saath seekhna jo har jawaab ko sahi ya galat mark kare.

The model sees inputs and their labels and learns to map one to the other, then predicts labels for new inputs. Classification and regression are supervised tasks. Most business ML is supervised.Model inputs aur unke labels dekhta hai aur ek se doosre tak pahunchna seekhta hai, phir naye inputs ke labels predict karta hai. Classification aur regression supervised tasks hain. Zyaadatar business ML supervised hai.

💡 Real exampleAsli example: Learning to predict default from past loans marked repaid or defaulted.Repaid ya defaulted mark kiye purane loans se default predict karna seekhna.

📖 Learn itSeekho: How machines learn: features, labels and trainingMachines kaise seekhti hain: features, labels aur training

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Unsupervised learning · Label · Classification

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: System message 🔊

🧸 Think of it likeAise samjho: Like the briefing a manager gives a new shop assistant on day one: who you are, how to talk to customers, and what never to do.Jaise manager pehle din naye shop assistant ko briefing deta hai: tum kaun ho, customers se kaise baat karni hai, aur kya kabhi nahi karna.

The system prompt defines who the assistant is, what it may and may not do, the tone and the output format. It applies to every user message. It's not a security boundary: sensitive rules must also be enforced in code.System prompt tay karta hai assistant kaun hai, kya kar sakta aur kya nahi, tone aur output format. Yeh har user message pe laagu hota hai. Yeh security boundary nahi hai: sensitive rules code mein bhi lagne chahiye.

💡 Real exampleAsli example: 'You are a polite support assistant for Acme Bank. Answer only from the provided policy.'System prompt: 'Aap Acme Bank ke vinamra support assistant ho. Sirf di gayi policy se jawaab do.'

📖 Learn itSeekho: Prompting: giving clear instructions to an LLMPrompting: LLM ko saaf instructions dena · Calling an LLM API: keys, requests and costsLLM API call karna: keys, requests aur costs

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Prompt · Guardrails

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: Top p, sampling 🔊

🧸 Think of it likeAise samjho: Like a creativity dial: low gives the safe, usual answer every time; high gives more surprising, varied answers.Creativity dial jaisa: kam pe har baar safe, aam jawaab; zyada pe zyada alag aur surprising jawaab.

Low temperature (near 0) makes the model pick the most likely tokens, so outputs are consistent, good for extraction and factual answers. Higher temperature adds variety for brainstorming and creative writing. Top p is a related setting that limits choices to the most likely tokens.Kam temperature (0 ke paas) pe model sabse likely tokens chunta hai, to outputs consistent, extraction aur factual jawaabon ke liye accha. Zyaada temperature brainstorming aur creative writing ke liye variety laata hai. Top p juda hua setting hai jo choices ko sabse likely tokens tak seemit karta hai.

💡 Real exampleAsli example: Temperature 0 for reading invoice fields; 0.9 for ten tagline ideas.Invoice fields padhne ke liye temperature 0; das tagline ideas ke liye 0.9.

📖 Learn itSeekho: What is an LLM? Tokens, context and temperatureLLM kya hai? Tokens, context aur temperature · Prompting: giving clear instructions to an LLMPrompting: LLM ko saaf instructions dena

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Softmax · Token · Prompt

Machine learningMachine learning Also calledIse yeh bhi kehte hain: Hold out set 🔊

🧸 Think of it likeAise samjho: Like the final exam paper kept sealed until the end, so nobody can practise on it.Final exam paper jaisa jo aakhir tak seal rehta hai, taaki koi uspe practice na kar sake.

Never train or tune on the test set; use it once for the final number. For time based data, the test set should be the most recent period. If you peek repeatedly, scores become too optimistic.Test set pe kabhi train ya tune mat karo; final number ke liye ek baar use karo. Time based data mein test set sabse haal ka period hona chahiye. Baar baar jhaankoge to scores zyaada optimistic ho jaate hain.

💡 Real exampleAsli example: Training on 2023 to 2025 data and testing on the first half of 2026.2023 se 2025 ke data pe training aur 2026 ki pehli chhamahi pe testing.

📖 Learn itSeekho: Train/test split and overfittingTrain/test split aur overfitting

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Validation set · Overfitting · Data leakage

Machine learningMachine learning Also calledIse yeh bhi kehte hain: Decision threshold 🔊

🧸 Think of it likeAise samjho: Like a pass mark: score above 40 and you pass, below and you don't. Moving the pass mark changes who gets through.Pass marks jaisa: 40 se upar pass, neeche fail. Pass marks badalne se badalta hai kaun aage jaata hai.

A model might say 0.37 chance of fraud. With a 0.5 threshold that's 'no'; with 0.3 it's 'yes'. Raising the threshold gives fewer false alarms (higher precision) but more misses (lower recall). Choose it from the business cost of each mistake, not by default.Model 0.37 fraud ka chance bata sakta hai. 0.5 threshold pe yeh 'nahi' hai; 0.3 pe 'haan'. Threshold badhaane se kam false alarms (zyaada precision) par zyaada miss (kam recall). Ise default se nahi, har galti ki business cost se chuno.

💡 Real exampleAsli example: Flagging payments above 0.3 so the review team catches 85% of fraud.0.3 se upar wali payments flag karna taaki review team 85% fraud pakde.

📖 Learn itSeekho: Measuring models: accuracy, precision and recallModels ko measure karna: accuracy, precision aur recall

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Precision · Recall · AUC (ROC AUC, PR AUC)

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: Tokenizer 🔊

🧸 Think of it likeAise samjho: Like the syllables of a word: an LLM reads text as small pieces, not whole words.Shabd ke akshar samooh jaisa: LLM text ko chhote tukdon mein padhta hai, poore shabdon mein nahi.

A tokenizer splits text into tokens; in English one token is about three quarters of a word, and Indian languages often need more tokens for the same meaning. Cost, speed and the context window all depend on token counts, so shorter prompts and outputs save money.Tokenizer text ko tokens mein todta hai; English mein ek token lagbhag teen chauthai shabd hai, aur Indian languages ko aksar same matlab ke liye zyaada tokens chahiye. Cost, speed aur context window sab token counts pe depend karte hain, isliye chhote prompts aur outputs paise bachaate hain.

💡 Real exampleAsli example: 'Unbelievable' may split into 'un', 'believ', 'able': three tokens.'Unbelievable' 'un', 'believ', 'able' mein toot sakta hai: teen tokens.

📖 Learn itSeekho: What is an LLM? Tokens, context and temperatureLLM kya hai? Tokens, context aur temperature · Calling an LLM API: keys, requests and costsLLM API call karna: keys, requests aur costs

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Context window · LLM (large language model)

Agents and toolsAgents aur tools Also calledIse yeh bhi kehte hain: Function calling, tool use 🔊

🧸 Think of it likeAise samjho: Like a receptionist who can't see the stock room but can call the store keeper and ask 'is item 4512 in stock?'Receptionist jaisa jo store room nahi dekh sakta par storekeeper ko phone karke pooch sakta hai 'item 4512 stock mein hai?'

You describe tools with names and JSON schemas; the model replies with a tool call and arguments; your code validates them, runs the tool and returns the result; the model continues. The model never runs code itself. Keep tools small and clear, check permissions and require confirmation for risky actions.Tools ko naam aur JSON schemas se describe karte ho; model tool call aur arguments ke saath jawaab deta hai; aapka code unhe validate karta, tool chalata aur result lautata hai; model aage badhta hai. Model khud kabhi code nahi chalata. Tools chhote aur saaf rakho, permissions check karo aur risky actions pe confirmation maango.

💡 Real exampleAsli example: The user asks 'where is order 4512?' and the model calls get_order_status(4512).User poochhta hai 'order 4512 kahan hai?' aur model get_order_status(4512) call karta hai.

📖 Learn itSeekho: Tool use (function calling): letting an LLM take actionsTool use (function calling): LLM ko action lene dena · AI agents: plan, act, observe, repeatAI agents: plan, act, observe, repeat

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Agent (AI agent) · MCP (Model Context Protocol) · Structured output

AI basicsAI basics Also calledIse yeh bhi kehte hain: Model training 🔊

🧸 Think of it likeAise samjho: Like practice sessions where the model looks at many examples and slowly gets better at the task.Practice sessions jaisa jahan model bahut saare examples dekh kar dheere dheere kaam mein behtar hota hai.

The algorithm makes predictions on training data, measures the loss, and adjusts parameters to reduce it, over many iterations. Training needs data, compute and evaluation on separate data. Pretraining a large model is very expensive; most teams fine tune or simply use existing models.Algorithm training data pe predictions karta hai, loss naapta hai, aur use kam karne ke liye parameters badalta hai, kai baar. Training ko data, compute aur alag data pe evaluation chahiye. Bade model ki pretraining bahut mehngi hai; zyaadatar teams fine tune karti hain ya seedhe existing models use karti hain.

💡 Real exampleAsli example: Training a churn model on 2 years of customer history takes 10 minutes on a laptop.2 saal ki customer history pe churn model training laptop pe 10 minute leti hai.

📖 Learn itSeekho: How machines learn: features, labels and trainingMachines kaise seekhti hain: features, labels aur training · Training vs inference: how ML runs in productionTraining vs inference: ML production mein kaise chalta hai

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Inference · Loss function · Epoch

MLOps and deploymentMLOps aur deployment 🔊

🧸 Think of it likeAise samjho: Like practising cricket with a tennis ball and then playing the match with a leather ball.Tennis ball se cricket practice karna aur phir match leather ball se khelna.

For example, a 30 day average in training versus a 7 day average in the live service. Prevent it with one shared feature pipeline or a feature store, logging live features and comparing them with training data.Jaise training mein 30 din ka average aur live service mein 7 din ka. Ek shared feature pipeline ya feature store, live features log karke training data se compare karne se ise roko.

💡 Real exampleAsli example: A model trained on prices in rupees served with prices in paise.Rupees mein prices pe train hua model jise paise mein prices di jaayein.

📖 Learn itSeekho: Training vs inference: how ML runs in productionTraining vs inference: ML production mein kaise chalta hai

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Feature store · Feature engineering

Deep learningDeep learning 🔊

🧸 Think of it likeAise samjho: Like a person who can ride a bicycle learning a scooter quickly: the balance is already learned.Jaise cycle chalaana jaanne wala scooter jaldi seekh leta hai: balance pehle se aata hai.

A model pretrained on millions of images or huge amounts of text already knows general patterns. Fine tuning it (or just its last layers) on a small dataset for your task works far better than training from scratch. It's the default approach in modern deep learning.Laakhon images ya bahut saare text pe pretrained model pehle se general patterns jaanta hai. Ise (ya sirf aakhri layers ko) apne task ke chhote dataset pe fine tune karna shuru se train karne se kahin behtar chalta hai. Aaj ki deep learning mein yahi default tarika hai.

💡 Real exampleAsli example: Adapting an image model to spot defects using 500 factory photos.500 factory photos se defects pakadne ke liye image model ko adapt karna.

📖 Learn itSeekho: Fine-tuning and LoRA: specialising a modelFine-tuning aur LoRA: model ko specialise karna

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Fine tuning · Foundation model

Deep learningDeep learning 🔊

🧸 Think of it likeAise samjho: Like a reader who looks at the whole sentence at once and links every word to the words that matter for it, instead of reading strictly one word at a time.Ek reader jaisa jo poora sentence ek saath dekhta hai aur har shabd ko unse jodta hai jo uske liye zaroori hain, ek ek shabd padhne ki jagah.

Transformers turn tokens into embeddings, then stack layers of self attention and feed forward networks with residual connections and normalisation. They process all tokens in parallel, which made training on huge datasets practical. They also power vision, speech and multimodal models.Transformers tokens ko embeddings mein badalte hain, phir self attention aur feed forward networks ki layers residual connections aur normalisation ke saath stack karte hain. Yeh saare tokens parallel mein process karte hain, jisse bade datasets pe training practical hui. Vision, speech aur multimodal models bhi inhi se chalte hain.

💡 Real exampleAsli example: GPT stands for Generative Pretrained Transformer.GPT ka matlab Generative Pretrained Transformer hai.

📖 Learn itSeekho: Transformers and attention: how LLMs readTransformers aur attention: LLMs kaise padhte hain

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Attention (self attention) · LLM (large language model) · Token

Machine learningMachine learning 🔊

🧸 Think of it likeAise samjho: Like a student who only read the chapter titles: they miss so much that even easy questions go wrong.Jaise student ne sirf chapter ke titles padhe: itna kuch chhoot gaya ki aasaan sawaal bhi galat.

Signs: low scores on both training and validation data, and errors that follow an obvious pattern the model ignores. It's the opposite of overfitting. Fixes: add better features, use a more flexible model (for example gradient boosting instead of a straight line), train longer, or reduce regularisation.Lakshan: training aur validation dono data pe kam score, aur aisi galtiyaan jinka saaf pattern model ignore karta hai. Yeh overfitting ka ulta hai. Upaay: behtar features jodo, zyaada flexible model lo (jaise seedhi line ke bajaye gradient boosting), zyaada der train karo, ya regularisation kam karo.

💡 Real exampleAsli example: Using a straight line to predict sales that clearly rise and fall with seasons.Mausam ke saath saaf upar neeche hone wali sales predict karne ke liye seedhi line use karna.

📖 Learn itSeekho: Train/test split and overfittingTrain/test split aur overfitting

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Overfitting · Bias variance trade off

Machine learningMachine learning 🔊

🧸 Think of it likeAise samjho: Like sorting a pile of photos into groups on your own, with nobody telling you what the groups should be.Photos ke dher ko khud groups mein baantna, bina kisi ke bataye ki groups kya hone chahiye.

Methods include clustering (customer segments), dimensionality reduction (simplifying many features) and anomaly detection (spotting unusual transactions). Results need human interpretation and are often used to explore data or create features.Tarike: clustering (customer segments), dimensionality reduction (bahut features ko simple karna) aur anomaly detection (ajeeb transactions pakadna). Results ko insaani samajh chahiye aur aksar data explore karne ya features banane mein use hote hain.

💡 Real exampleAsli example: Spotting a strange login pattern no one labelled as an attack yet.Aisa ajeeb login pattern pakadna jise abhi kisi ne attack label nahi kiya.

📖 Learn itSeekho: How machines learn: features, labels and trainingMachines kaise seekhti hain: features, labels aur training

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Clustering · Supervised learning

Machine learningMachine learning Also calledIse yeh bhi kehte hain: Dev set 🔊

🧸 Think of it likeAise samjho: Like mock tests during the year: you use them to decide how to study, not to get your final marks.Saal bhar ke mock tests jaisa: inse tay karte ho kaise padhna hai, final marks ke liye nahi.

You train on the training set, compare models and hyperparameters on the validation set, and keep the test set for the final check. Cross validation rotates the validation part for a more reliable estimate.Training set pe train karo, validation set pe models aur hyperparameters compare karo, aur final check ke liye test set rakho. Cross validation zyaada reliable estimate ke liye validation hissa baari baari badalta hai.

💡 Real exampleAsli example: Choosing between tree depth 4 and 8 based on validation F1.Validation F1 ke hisaab se tree depth 4 aur 8 mein chunna.

📖 Learn itSeekho: Train/test split and overfittingTrain/test split aur overfitting

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Test set · Cross validation · Hyperparameter

AI basicsAI basics 🔊

🧸 Think of it likeAise samjho: Like describing a mango with numbers: sweetness 8, size 6, ripeness 9. That list of numbers is a vector.Aam ko numbers se describe karna: meethaapan 8, size 6, pakaapan 9. Numbers ki yeh list ek vector hai.

Models work on vectors and matrices. A house can be [3, 1200, 8] for bedrooms, area and distance; an embedding is a long vector that captures meaning. Operations like the dot product and cosine similarity compare vectors.Models vectors aur matrices pe kaam karte hain. Ghar [3, 1200, 8] ho sakta hai, bedrooms, area aur doori ke liye; embedding ek lamba vector hai jo matlab pakadta hai. Dot product aur cosine similarity jaise operations vectors compare karte hain.

💡 Real exampleAsli example: A word embedding with 768 numbers.768 numbers wala word embedding.

📖 Learn itSeekho: Vectors and matrices: the language of AIVectors aur matrices: AI ki bhasha

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Embedding · Cosine similarity

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: Vector store, pgvector 🔊

🧸 Think of it likeAise samjho: Like a library arranged by meaning instead of title, so books about the same idea sit on the same shelf.Title ki jagah matlab ke hisaab se lagi library jaisa, taaki ek hi idea wali kitaabein ek shelf pe hon.

It stores vectors with their text and metadata and uses approximate nearest neighbour indexes to search millions quickly, with filters for permissions, dates or departments. Options include pgvector in PostgreSQL and dedicated vector databases. It's the storage layer for RAG.Yeh vectors ko unke text aur metadata ke saath rakhta hai aur approximate nearest neighbour indexes se laakhon mein jaldi search karta hai, permissions, dates ya departments ke filters ke saath. Options: PostgreSQL mein pgvector aur dedicated vector databases. Yeh RAG ki storage layer hai.

💡 Real exampleAsli example: Storing 25,000 policy chunks and finding the 5 most relevant in 30 ms.25,000 policy chunks store karna aur 30 ms mein 5 sabse relevant dhoondhna.

📖 Learn itSeekho: Embeddings and vector databases: searching by meaningEmbeddings aur vector databases: matlab se search karna · RAG: letting an LLM answer from your documentsRAG: LLM ko aapke documents se jawaab dene dena

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Embedding · RAG (retrieval augmented generation) · Semantic search

Agents and toolsAgents aur tools Also calledIse yeh bhi kehte hain: Agentic workflow, pipeline 🔊

🧸 Think of it likeAise samjho: Like a fixed assembly line where one station uses AI, versus an agent, which is more like a worker deciding the steps on their own.Ek fixed assembly line jaisa jismein ek station AI use karta hai, jabki agent ek worker jaisa hai jo steps khud tay karta hai.

Many business processes have known steps: classify the ticket, fetch context, draft a reply, check policy. A workflow hard codes the order and uses the LLM only where judgement is needed. It's cheaper, faster and easier to test than an autonomous agent; add agent style decisions only where they clearly help.Kai business processes ke steps pata hote hain: ticket classify karo, context lao, reply draft karo, policy check karo. Workflow order ko code mein tay karta hai aur LLM sirf wahan use karta hai jahan samajh chahiye. Yeh autonomous agent se sasta, tez aur test karne mein aasaan hai; agent jaise faisle wahin jodo jahan saaf faayda ho.

💡 Real exampleAsli example: Invoice processing: OCR, LLM extracts fields, code validates totals, human reviews exceptions.Invoice processing: OCR, LLM fields nikaalta hai, code totals validate karta hai, insaan exceptions dekhta hai.

📖 Learn itSeekho: AI agents: plan, act, observe, repeatAI agents: plan, act, observe, repeat

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Agent (AI agent) · Tool calling

LLMs and GenAILLMs aur GenAI Also calledIse yeh bhi kehte hain: Zero shot prompting 🔊

🧸 Think of it likeAise samjho: Like asking a new helper to 'make tea for five' without showing them how: they rely on what they already know.Naye helper ko bina dikhaye 'paanch logon ki chai bana do' kehna: woh pehle se jo jaanta hai uspe chalta hai.

Modern LLMs handle many tasks zero shot: 'Classify this ticket as billing, delivery or technical.' If results are inconsistent, add a few examples (few shot) or a clearer output format.Aaj ke LLMs kai kaam zero shot kar lete hain: 'Is ticket ko billing, delivery ya technical mein classify karo.' Results consistent na hon to kuch examples (few shot) ya saaf output format jodo.

💡 Real exampleAsli example: Asking for a summary of a meeting transcript with no sample summaries.Bina sample summaries ke meeting transcript ka summary maangna.

📖 Learn itSeekho: Prompting: giving clear instructions to an LLMPrompting: LLM ko saaf instructions dena

Open the full pagePoora page kholo →

🔗 RelatedJude hue shabd: Few shot prompting · Prompt