Machine learning engineer interview questions in English

1. Design a system that recommends videos on a home page.

ML system design

What they listen for: Clarification, a two-stage design, metrics, and serving constraints.

“First: what's the goal, watch time or satisfaction, and what's the latency budget? I'd use two stages: candidate generation to get a few hundred videos quickly, then a ranking model to order them. Offline I'd evaluate with ranking metrics; online with an A/B test on watch time and a satisfaction guardrail. Then serving, monitoring and retraining.”

2. How do you know a model is ready for production?

ML fundamentals

What they listen for: Offline metrics, comparison to baseline, online testing and operational readiness.

“Offline, it should beat the current model or baseline on the agreed metric, on a properly held-out test set. Then operationally: it meets the latency and memory budget, and we have monitoring and a rollback plan. Finally, an online A/B test, because offline gains don't always translate.”

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

ML fundamentals

What they listen for: Distinguishing input drift from concept drift, and practical monitoring.

“Data drift is when the input data in production changes from what the model was trained on. Concept drift is when the relationship between inputs and the target changes. I monitor feature distributions against the training data, and I monitor prediction quality when labels arrive. Alerts trigger an investigation or retraining.”

4. Your model is too slow to serve. What can you do?

ML system design

What they listen for: Several options with trade-offs.

“A few options. Make the model smaller, through distillation or quantisation. Cache predictions for common inputs. Precompute results offline where freshness allows. Or batch requests on the server. Each costs something, usually a bit of accuracy or freshness, so I'd choose based on the latency budget.”

5. Tell me about a model that performed worse in production than in testing.

Behavioural

What they listen for: Honest diagnosis, often leakage or training-serving skew, and the fix.

“Our churn model had great offline results but did poorly live. I found that one feature was calculated differently in training, from a cleaned batch table, than in serving, from raw events. That's training-serving skew. We moved feature logic into a shared feature store so both paths used the same code.”

6. How do you handle an imbalanced dataset?

ML fundamentals

What they listen for: Right metrics first, then techniques.

“First, the right metrics: precision, recall or PR-AUC instead of accuracy. Then techniques: class weights, resampling, or adjusting the decision threshold to the business cost of each error. Often just tuning the threshold is the biggest win.”

7. Implement a function to compute precision and recall.

Coding

What they listen for: Correct definitions, clean code and edge cases.

“I'll count true positives, false positives and false negatives in one pass over the predictions and labels. Precision is true positives over predicted positives, recall is true positives over actual positives. One edge case: if there are no predicted positives, precision divides by zero, so I'll return zero or None. Which do you prefer?”

8. How would you build a system to detect fraudulent transactions?

ML system design

What they listen for: Real-time constraints, label delay, cost of errors and adversaries.

“Questions first: how fast must we decide, and what does a false positive cost? Probably a real-time score within about 100 milliseconds, with features like transaction velocity and device history. Labels come late, through chargebacks, so I'd account for that delay. Fraudsters adapt, so retraining and rules for new patterns are important.”

9. When would you use an LLM versus a smaller specialised model?

ML fundamentals

What they listen for: Cost, latency, data and quality trade-offs, not hype.

“An LLM is great when the task is broad and language-heavy and you have little labelled data, or for prototyping fast. A smaller model is usually better when the task is narrow, you have good labelled data, and you need low latency or low cost at high volume. I'd often prototype with an LLM and move to a smaller model if volume grows.”

10. How do you work with data scientists?

Behavioural

What they listen for: Clear responsibilities and shared standards.

“We agree early on what production needs: latency, input format, how features are computed. Data scientists focus on modelling, I focus on making it reliable in production, but we share code and reviews. The goal is that a model can move from notebook to production without being rewritten.”

What questions are asked in a machine learning engineer interview?

Coding, ML fundamentals such as evaluation and imbalance, ML system design for recommendation, search or fraud, production topics like serving, latency and drift, and behavioural questions about shipping models with other teams.

How is an ML engineer interview different from a data scientist interview?

ML engineer loops put more weight on coding, system design and production reliability. Data scientist loops put more weight on statistics, experimentation and product analysis. Both ask about model evaluation.

How do I explain an ML system design out loud in English?

Signpost first: 'I'll cover data, features, model, evaluation, serving and monitoring.' Then go through each in two or three sentences and ask the interviewer where to go deeper. The signpost keeps both of you on track.