Question 1

Tell me about yourself.

Accepted Answer

Use present, past, future, in about 60 to 90 seconds. Present: "I'm a Java backend developer with four years at a fintech company, where I build payment APIs." Past: "Over the last year I moved towards AI. I built a document question answering app with RAG and an evaluation set, and I used an LLM to classify support tickets at work, which cut manual triage by about 30%." Future: "Now I want to do this full time as an AI engineer, and this role fits because you build LLM features on top of real backend systems, which is exactly where my strengths meet."

Question 2

Why do you want to move into AI, and why this particular role?

Accepted Answer

Give a real reason and connect it to evidence. "In my testing work I kept seeing the same problem: thousands of test cases written and checked by hand. When I tried an LLM to draft test cases, it saved hours, but it also made mistakes I had to catch. That got me interested in building AI systems that are actually reliable. I've since completed projects in RAG and evaluation. I'm applying for this role because it focuses on shipping AI features to users, and my background in quality and testing directly helps with evaluating AI output."

Question 3

Why do you want to work here?

Accepted Answer

Show you've done homework. Mention one thing about their product, one thing about how they work, and link both to you. "I've used your app for expense reports, and I noticed the receipt scanning recently got much better, so I read your engineering blog about it. I like that your team publishes how you evaluate models, because I care about measuring AI quality. And since you serve small businesses in India, my experience building for that market would help from day one."

Question 4

Explain supervised, unsupervised and reinforcement learning, with an example of each.

Accepted Answer

Supervised learning learns from examples that have the right answer, called labels. For example, emails marked spam or not spam, so the model learns to predict spam. Unsupervised learning has no labels; it finds structure by itself, like grouping customers into segments by behaviour. Reinforcement learning learns by trial and error with rewards, like a game playing agent or a robot that gets a reward when it does the task well. LLMs also use reinforcement learning from human feedback to make answers more helpful.

Question 5

What is overfitting, and how do you prevent it?

Accepted Answer

Overfitting is when a model memorises the training data, including its noise, so it scores very well on training data but badly on new data. It's like a student who memorises last year's answers but can't solve a new question. You spot it when training accuracy is high and validation accuracy is much lower. To prevent it: get more data, use a simpler model, add regularisation, use cross validation, stop training early, and for neural networks use dropout.

Question 6

What is the bias variance trade off?

Accepted Answer

Bias is error from a model that's too simple, so it misses real patterns. That's underfitting. Variance is error from a model that's too sensitive to the training data, so small changes in the data change it a lot. That's overfitting. Making a model more complex usually lowers bias but raises variance. The goal is the sweet spot where total error on new data is lowest, which we find using validation data. For example, a straight line through curved data has high bias; a very deep decision tree has high variance.

Question 7

Why do we split data into training, validation and test sets?

Accepted Answer

The training set is what the model learns from. The validation set is used while building, to compare models and tune settings like depth or learning rate. The test set is kept locked away until the very end to get an honest estimate of how the model will do on new data. If you tune on the test set, you slowly overfit to it and your final number becomes too optimistic. For time based data, split by time, so you train on the past and test on the future.

Question 8

What is data leakage? Give an example.

Accepted Answer

Leakage is when information that won't be available at prediction time sneaks into training, so the model looks great in testing and then fails in production. A classic example: predicting whether a customer will cancel, and including a feature like "refund issued", which only happens after they cancel. Another is scaling features using statistics from the whole dataset before splitting, so the test set influences training. To avoid it, ask for each feature: would I know this value at the moment I make the prediction? And do all preprocessing inside the training split.

Question 9

Explain precision and recall. When would you prefer one over the other?

Accepted Answer

Precision: of everything the model flagged as positive, how many were really positive. Recall: of all the real positives, how many the model caught. For a cancer screening test or fraud detection, missing a real case is very costly, so we prefer high recall, even if we get some false alarms. For a spam filter, putting a real email into spam is annoying, so we prefer high precision. F1 score combines both when you need one number, and you can move between them by changing the decision threshold.

Question 10

Your model is 99% accurate, but the business says it's useless. How is that possible?

Accepted Answer

Probably the classes are imbalanced. If only 1% of transactions are fraud, a model that always says "not fraud" is 99% accurate and catches zero fraud. Accuracy hides that. I'd look at the confusion matrix, precision and recall for the rare class, and metrics like PR AUC. Then I'd check the business goal: maybe they need to catch 80% of fraud while keeping false alarms low enough for the review team to handle, and I'd tune the threshold for that.

Question 11

What is regularisation? What's the difference between L1 and L2?

Accepted Answer

Regularisation adds a penalty for large weights to the loss, which keeps the model simpler and reduces overfitting. L2, also called ridge, penalises the square of weights, so it shrinks all weights smoothly towards zero but rarely to exactly zero. L1, also called lasso, penalises the absolute value, which pushes some weights to exactly zero, so it also does feature selection. In neural networks we also use dropout and weight decay, which is L2 in practice.

Question 12

Explain gradient descent in simple words. What does the learning rate do?

Accepted Answer

Imagine you're on a hill in fog and want to reach the bottom. You feel which way the ground slopes and take a step downhill, again and again. Gradient descent does the same with the model's error: the gradient tells us which direction increases the error, so we move the weights a small step in the opposite direction. The learning rate is the step size. Too big and you overshoot and bounce around; too small and training is very slow. In practice we use mini batches and optimisers like Adam.

Question 13

What is cross validation and when would you not use standard k fold?

Accepted Answer

K fold cross validation splits training data into k parts, trains on k minus 1 and validates on the remaining part, rotating through all parts. Averaging gives a more reliable estimate than a single split, which matters with small data. I wouldn't use plain k fold for time series, because it lets the model see the future; I'd use time based splits instead. And if one user has many rows, I'd use group k fold so the same user isn't in both training and validation.

Question 14

How would you implement logistic regression from scratch?

Accepted Answer

I'd explain the steps, then code them in NumPy. One, initialise weights and bias to zeros. Two, compute z = X·w + b and the prediction p = sigmoid(z). Three, the loss is binary cross entropy. Four, the gradient is simple: dw = Xᵀ(p − y) / n and db = mean(p − y). Five, update w and b by subtracting learning rate times the gradient, and loop for some epochs. I'd add a check that the loss goes down, clip probabilities to avoid log of zero, and test on a tiny dataset where I know the answer.

Question 15

Explain self attention. Why do we divide by the square root of d_k?

Accepted Answer

Each token creates three vectors: a query, "what am I looking for?", a key, "what do I contain?", and a value, "what do I pass on?". For each token, we compare its query with every key using a dot product, which gives relevance scores. Softmax turns the scores into weights, and the output is a weighted sum of the values. So each token builds a new representation by looking at the others. We divide by the square root of d_k because dot products grow with dimension; without scaling, softmax becomes very peaked and its gradients become tiny, which slows training.

Question 16

What's the computational cost of self attention, and how can it be reduced?

Accepted Answer

For a sequence of n tokens with dimension d, attention compares every token with every other token, so time is roughly O(n² · d) and the attention matrix needs O(n²) memory. That's why long contexts are expensive. Approaches to reduce it include FlashAttention, which computes exact attention in tiles to save memory and is much faster on GPUs; sparse or sliding window attention, which only looks at nearby tokens; linear attention approximations; and at inference, the KV cache plus grouped query attention to reduce memory per token.

Question 17

What is the KV cache in LLM inference?

Accepted Answer

When an LLM generates text one token at a time, each new token needs attention over all previous tokens. Without a cache, we'd recompute the keys and values of every previous token at every step, which is wasteful. The KV cache stores the keys and values computed so far, so each step only computes them for the new token. That makes generation much faster, but the cache uses a lot of GPU memory, growing with sequence length and batch size, which is why techniques like grouped query attention and paged attention, as in vLLM, matter.

Question 18

What's the gradient of softmax combined with cross entropy loss?

Accepted Answer

If p is the softmax output and y is the one hot true label, the gradient of the cross entropy loss with respect to the logits is simply p − y. Intuitively, for the correct class the gradient is p − 1, which is negative, so training pushes that logit up; for wrong classes it's p, pushing those logits down in proportion to how confident the model wrongly was. This clean result is why softmax and cross entropy are combined in one function in frameworks, which is also more numerically stable using the log sum exp trick.

Question 19

Walk me through implementing multi head attention with a causal mask.

Accepted Answer

Input x has shape batch × sequence × d_model. One, project x with three linear layers into Q, K and V. Two, reshape each to batch × heads × sequence × head_dim, where head_dim = d_model / heads. Three, scores = Q·Kᵀ / √head_dim, giving batch × heads × seq × seq. Four, for a causal mask, set scores above the diagonal to minus infinity, so a token can't see the future. Five, softmax over the last dimension, optional dropout, then multiply by V. Six, merge the heads back to batch × sequence × d_model and apply an output projection. I'd test shapes at each step and check that changing a future token doesn't change earlier outputs.

Question 20

What is LoRA, and why is it popular?

Accepted Answer

LoRA, low rank adaptation, fine tunes a big model by freezing the original weights and learning a small update for some weight matrices, written as the product of two thin matrices, A and B. Because the rank is small, you train maybe less than 1% of the parameters, which needs far less GPU memory and storage. You can keep one base model and swap small LoRA adapters for different tasks. QLoRA goes further by keeping the base model in 4 bit precision, so you can fine tune large models on a single GPU.

Question 21

What are vanishing gradients, and what fixed them in modern networks?

Accepted Answer

In deep networks, gradients are multiplied through many layers during backpropagation. If those factors are below one, the gradient shrinks to nearly zero by the time it reaches early layers, so they stop learning. Fixes that made deep networks work: activations like ReLU instead of sigmoid; residual or skip connections, which give gradients a direct path back; normalisation layers like LayerNorm; careful weight initialisation; and optimisers like Adam. Transformers rely heavily on residual connections and LayerNorm.

Question 22

How do you reproduce a paper, and what do you do when your results don't match?

Accepted Answer

I read the paper for its core claim and the exact setup: data, preprocessing, model size, hyperparameters and evaluation. If code exists, I run it first to get a reference. Then I implement step by step, testing each part on tiny data, and match the simplest experiment before the full one. If results don't match, I check the usual suspects: evaluation differences, data preprocessing, random seeds and variance across runs, learning rate schedule, and details hidden in the appendix. I log every experiment, and if a gap remains, I write it up honestly; that's useful research too.

Question 23

What is an ablation study and why does it matter?

Accepted Answer

An ablation study removes or changes one part of a system at a time to see how much each part contributes. If my new model has three changes and improves accuracy by 4 points, ablations tell me whether all three matter or one change does all the work. It matters because it turns "it works" into "we know why it works", prevents unnecessary complexity, and is expected in any credible paper. Good ablations change one thing at a time and use several random seeds, so differences aren't just noise.

Question 24

How do you make model training reproducible?

Accepted Answer

Version everything that affects the result: the code with git, the data with a snapshot or tool like DVC, the environment with a Docker image and pinned packages, and the configuration and hyperparameters in a tracked config file. Set random seeds. Log every run, with its code commit, data version, parameters, metrics and the resulting model, in an experiment tracker like MLflow. Then anyone can rerun or explain a model months later, which also matters for audits.

Question 25

How do you serve an open source LLM efficiently on GPUs?

Accepted Answer

Use an inference server built for LLMs, like vLLM or TGI, rather than a plain web server. They provide continuous batching, which adds new requests into the running batch so the GPU stays busy, and paged attention to manage KV cache memory. Choose the smallest model that meets quality, and consider quantisation, like 8 bit or 4 bit, to fit on cheaper GPUs. Then autoscale on queue length or tokens per second, set max token limits, and monitor time to first token, tokens per second, GPU memory and cost per request.

Question 26

Design a platform to serve an LLM to 1,000 requests per second.

Accepted Answer

Clarify model size, prompt and output lengths, and latency targets, especially time to first token. Requests go through a gateway with authentication, rate limits per customer and routing. Behind it, a pool of GPU servers running an inference engine like vLLM with continuous batching and paged KV cache, autoscaled on queue depth. Add a prefix and response cache for repeated prompts, and route simple requests to a smaller model. Estimate capacity from tokens per second per GPU, plan headroom for peaks, and spread across zones. Monitor time to first token, tokens per second, queue time, errors and cost per thousand tokens, and roll out new model versions with canaries.

Question 27

Walk me through a project you're proud of.

Accepted Answer

Use STAR and go deep on your own decisions. "Situation: our support team was overwhelmed with repeated questions. Task: I wanted to build an assistant that answers from our help docs. Action: I built a RAG pipeline, but the first version answered only 60% of test questions correctly. I created an evaluation set of 80 real questions, found that most failures came from chunks splitting tables, switched to structure aware chunking and added reranking. Result: correct answers rose to 86%, and in a pilot it resolved about a quarter of tickets without an agent." Be ready for "why did you choose that?" at every step.

Question 28

Tell me about a time something you built failed.

Accepted Answer

Pick a real failure with a lesson, not a fake one. "I shipped a classification model that looked great offline, with 92% accuracy, but within two weeks users complained. I had split data randomly, and the same customers appeared in both training and test, so the score was too optimistic. I owned it with my manager, re split by customer, got a realistic 81%, and improved features to reach 86%. Since then, I always ask how the data is grouped before splitting, and I add a review of the evaluation setup to my checklist."

Question 29

Tell me about a disagreement with a teammate or stakeholder.

Accepted Answer

Show you disagree with data and respect. "A product manager wanted to launch our AI reply feature to all customers. I was worried because our tests showed it sometimes made up refund policies. Instead of just saying no, I showed five real examples and the error rate from our eval set, and proposed launching to 10% of tickets with agent approval first. We agreed on criteria to expand. Two weeks later the error rate was below our threshold and we expanded. We kept a good relationship because we focused on the shared goal."

Question 30

Explain a technical concept to me as if I'm not technical.

Accepted Answer

Pick something from your work and use an everyday analogy. For example, explaining RAG: "Imagine a very smart new employee who has read a lot but knows nothing about our company. If you ask them a policy question, they might confidently guess. RAG is like giving them the right pages of our policy manual before they answer, and asking them to point to the page. So answers are based on our actual documents, and you can check them." Then check understanding: "Does that make sense?"

Question 31

How do you keep up with the fast changes in AI?

Accepted Answer

Be specific and show you filter noise. "I follow a few high quality sources instead of everything: release notes from the main model providers, a couple of newsletters, and papers when something important comes up. But mostly I learn by building: when something new appears, like a new tool calling feature, I try it on a small project within a week and write a short note about what worked. I also review what changed in our own evals when we upgrade models, which teaches me more than reading headlines."

Question 32

How do you use AI coding assistants in your work, and how do you make sure the code is right?

Accepted Answer

Be honest and responsible. "I use AI assistants daily for boilerplate, tests, unfamiliar libraries and explaining code. But I treat the output like code from a new colleague: I read every line before accepting it, I make sure there are tests, I check security sensitive parts like authentication and SQL myself, and I never paste secrets or confidential data into tools that aren't approved. It makes me faster, but I'm still responsible for what I ship."

Question 33

Do you have any questions for us?

Accepted Answer

Always say yes and ask two or three real questions. Good ones: "What would success look like for this role in the first 6 months?", "What's the biggest challenge the team is facing with AI right now?", "How do you evaluate model or feature quality before shipping?", "How is the team split between building new features and maintaining existing ones?", and "What do people who do well here have in common?". Avoid asking about leave or salary in a technical round, and don't ask things that are on their website.

Question 34

What are your salary expectations?

Accepted Answer

Research first, using sites like Glassdoor, AmbitionBox, levels.fyi and people in similar roles, then give a range based on the role, not on your current pay. "Based on my research for this role and level in Bangalore, and my experience, I'm looking at a range of ₹22 to ₹28 lakh fixed. I'm flexible depending on the overall package, and I'm most interested in the role itself." If they insist on your current salary, you can share it, but steer back to the market range for the new role. Never accept or reject on the spot; ask for the offer in writing.

Question 35

You need to check whether 1 million IDs are in a collection of 1 million known IDs. What's the efficient approach?

Accepted Answer

Put the known IDs in a set and use `in`. Membership checks in a set are O(1) on average, but O(n) in a list. With a list this becomes about a trillion comparisons.

Question 36

What does df.groupby('city')['amount'].mean() return?

Accepted Answer

One average amount per city. groupby splits the rows by city, then mean() averages 'amount' within each group, giving a Series indexed by city.

Question 37

Salaries in a team are mostly around ₹8 LPA, with one executive at ₹2 crore. Which best describes a typical salary?

Accepted Answer

The median. One extreme value pulls the mean up a lot. The median, the middle value, isn't affected by outliers.

Question 38

Users who turned on a new 'focus mode' spend 30% more time in the app. Can you say focus mode caused it?

Accepted Answer

No. People who opt in may already be more engaged, so run a randomised A/B test. This is self-selection: the opt-in group differs from everyone else. Only random assignment isolates the feature's effect.

Question 39

What does a 95% confidence interval for a model's accuracy mean?

Accepted Answer

If we repeated the evaluation on many fresh samples, about 95% of the intervals built this way would contain the true accuracy. A confidence interval describes the procedure's long-run coverage of the true value, not a per-prediction probability.

Question 40

Training accuracy is 98% and test accuracy is 65%. What's happening, and what's a reasonable fix?

Accepted Answer

Overfitting, so simplify the model, regularise or get more data. A large train/test gap means the model memorised the training data. Reduce its complexity, regularise, or add data.

Question 41

A fraud model is 99% accurate, and 1% of transactions are fraud. What's your first reaction?

Accepted Answer

Compare against the baseline and check precision and recall, because predicting 'not fraud' for everything also scores 99%. With rare classes, accuracy is misleading. Precision and recall show whether fraud is actually being caught.

Question 42

Why use k-fold cross-validation?

Accepted Answer

It gives a more reliable performance estimate by averaging over several train/validation splits. One split can be lucky or unlucky. Averaging across folds reduces that variance.

Question 43

Training loss jumps around wildly and never settles. What do you try first?

Accepted Answer

Decrease the learning rate. Steps that are too large overshoot the minimum. A smaller learning rate, or a warm-up schedule, stabilises training.

Question 44

What does self-attention compute in a transformer?

Accepted Answer

For each token, how much to draw on every other token, giving context-aware representations. Queries and keys produce attention weights, and the values are mixed by those weights, so each token's meaning depends on its context.

Question 45

LoRA reduces fine-tuning cost mainly by…

Accepted Answer

Freezing the base weights and training small low-rank adapter matrices. Only a tiny fraction of parameters are trained, which cuts memory and compute while keeping most of the quality.

Question 46

Why keep a validation set separate from the test set?

Accepted Answer

To tune hyperparameters without biasing the final test estimate. If you tune against the test set, it stops being an unbiased estimate of real-world performance.

Question 47

Vanishing gradients in deep networks are most associated with…

Accepted Answer

Saturating activations like sigmoid in many stacked layers, and are eased by ReLU, residual connections and normalisation. Repeatedly multiplying small derivatives shrinks gradients towards zero. Modern architectures add shortcuts and normalisation to fix this.

Question 48

Tell me about yourself and why you're moving into this AI role.

Accepted Answer

I'm a backend engineer with 8 years in fintech, building payment APIs that handle about 2 million requests a day. Over the last six months I've been moving into LLM engineering: I built a support-triage agent with tool calling and an eval suite, which raised correct routing from 71% to 89% on 200 real tickets. This role fits because it combines what I'm already good at, reliable APIs and production debugging, with the agent work I enjoy most. In the first months I'd bring production discipline: tracing, evals and cost controls for your agents.

Question 49

Walk me through an AI or ML project you built end to end.

Accepted Answer

I built a policy assistant for branch staff who spent about 25 minutes a day searching PDFs. I indexed 140 policy documents with chunking and metadata, used hybrid search, and required citations with a 'not found' rule. I chose RAG over fine-tuning because the policies change monthly. On a 60-question eval set built with staff, correct-and-cited answers went from 68% to 90% after I added re-ranking. It's deployed as an internal web app. The main failure was tables in PDFs; next I'd add table-aware parsing.

Question 50

Tell me about a time something you built failed. What did you do?

Accepted Answer

Our nightly data pipeline silently loaded duplicate orders for three days, inflating revenue on a dashboard by 12%. I owned that pipeline. I traced it to a retry that re-appended a partial load, so the job wasn't idempotent. I fixed the data, switched the load to overwrite-by-partition, and added a uniqueness check that fails the run. Since then we've had no duplicate incidents, and I now design every pipeline to be safely re-runnable.

Question 51

Design a fraud-detection system for UPI payments.

Accepted Answer

I'd start from requirements: decisions in under 100 ms at peak volume, and an explicit cost for missed fraud vs blocking a genuine payment. Features combine real-time signals, such as amount versus the user's usual, transactions in the last 10 minutes and a new device or payee, with historical aggregates from a feature store, all computed only from data available at payment time. A gradient-boosted model fits tabular signals well. I'd set the threshold to minimise expected cost and send a middle band to manual review. It would be served behind a low-latency API with streaming feature updates. I'd monitor input drift and precision and recall as chargeback labels arrive weeks later, and retrain on a schedule.

Question 52

Explain how a transformer works to an engineer, and how you'd reduce its training cost.

Accepted Answer

Text is tokenised and each token mapped to a vector, with positional information added. In each layer, self-attention projects tokens into queries, keys and values; the attention weights from queries and keys mix the values, so each token's representation reflects its context. Multiple heads capture different relationships, followed by a feed-forward network, with residual connections and normalisation for stable training. The model is trained to predict the next token. To cut cost I'd use mixed precision, efficient attention kernels, a cleaner and deduplicated dataset, and parameter-efficient fine-tuning instead of full training where possible.

AI Research Engineer interview questions and answers

How an AI Research Engineer interview usually goes

👋 Opening: about you

🧱 Fundamentals

🛠️ Role skills

🏗️ Design and case questions

🤝 Projects and behaviour

❓ Closing: your questions and salary

🧪 Quick quiz: test yourself

Python & coding

Statistics & data

ML fundamentals

Deep learning

Communication

ML system design

More questions to prepare

Practise with a score and gap analysis

Also for AI Research Engineer

Sources