Question 1

How would you deploy a model to production?

Accepted Answer

Package the model with its preprocessing, serve it behind an API (or run batch jobs), version the model and data, test against the current model, roll out gradually and monitor.

Question 2

Batch vs real-time inference - how do you choose?

Accepted Answer

Batch when predictions can be computed ahead (daily recommendations) - cheaper and simpler. Real-time when they depend on fresh input (fraud at checkout) - needs low latency and scaling.

Question 3

What is data drift, and how do you detect it?

Accepted Answer

The input data distribution changes from training, so performance degrades. Monitor feature distributions (for example with PSI), prediction distributions and, when labels arrive, real accuracy; retrain on triggers.

Question 4

What is training-serving skew?

Accepted Answer

Features computed differently in training and serving - different code, timing or defaults. Fix it with shared feature code or a feature store, and log serving features to compare.

Question 5

Explain the transformer architecture briefly.

Accepted Answer

Self-attention lets each token weigh every other token in the sequence, processed in parallel, with positional information added. Stacks of attention and feed-forward layers form encoders and decoders.

Question 6

What is RAG, and when would you use it instead of fine-tuning?

Accepted Answer

Retrieval-augmented generation fetches relevant documents (via embeddings and vector search) and gives them to the LLM as context. Use it for changing or private knowledge; fine-tune to change style or behaviour.

Question 7

How do you evaluate an LLM feature?

Accepted Answer

A fixed test set of real questions with expected answers, automatic checks (format, groundedness), human or model-graded scoring, plus monitoring of cost, latency and user feedback in production.

Question 8

How would you reduce inference latency?

Accepted Answer

Smaller or distilled models, quantisation, batching requests, caching repeated results, GPU serving where justified, and moving work out of the request path.

Question 9

Design a recommendation system for an e-commerce home page.

Accepted Answer

Candidate generation (collaborative filtering, embeddings) then ranking with a model on user and item features, with business rules on top. Cover data pipelines, cold start, offline metrics and an A/B test.

Question 10

Tell me about a model you put into production.

Accepted Answer

The problem, the model, how it was served, what you monitored and one production issue you fixed.

Question 11

What is the difference between a data scientist and an ML engineer?

Accepted Answer

Data scientists focus on analysis and building models; ML engineers focus on deploying, scaling and maintaining models in production. The roles overlap at many companies.

Question 12

Are LLM questions asked in ML engineer interviews now?

Accepted Answer

Increasingly - RAG, embeddings, evaluation and prompt design come up often, especially for AI product teams.

Question 13

What coding skills do ML engineers need?

Accepted Answer

Strong Python, data structures and algorithms, SQL, and familiarity with tools like Docker and a cloud platform.

Machine learning engineer interview questions and answers

What to revise before the interview

1.How would you deploy a model to production?

2.Batch vs real-time inference - how do you choose?

3.What is data drift, and how do you detect it?

4.What is training-serving skew?

5.Explain the transformer architecture briefly.

6.What is RAG, and when would you use it instead of fine-tuning?

7.How do you evaluate an LLM feature?

8.How would you reduce inference latency?

9.Design a recommendation system for an e-commerce home page.

10.Tell me about a model you put into production.

Frequently asked questions

Company interview guides for this role

Related roles and topics

Practise these questions for your next interview