Skip to content
Quantoros

Guides

Machine learning interview preparation: questions and a plan

· 10 min read

ML engineer interviews combine coding, ML theory, system design and a discussion of your projects. This guide lists the questions that come up again and again, with short answers, and a plan that fits around a full-time job.

Typical stages

  1. Recruiter screen and a short technical conversation.
  2. Coding: Python, data structures and data manipulation; SQL for many roles.
  3. ML theory and practical modelling questions.
  4. ML system design: design a recommender, a fraud detector or a search ranker.
  5. Take-home or project deep dive.
  6. Behavioural interview.

ML theory: questions with short answers

  • Bias–variance trade-off: simple models underfit (high bias), flexible ones overfit (high variance); regularisation, more data and validation choose the balance.
  • L1 vs L2 regularisation: L1 drives coefficients to zero and selects features; L2 shrinks them smoothly and handles correlated features better.
  • How to choose a metric: start from the business cost of errors; for imbalanced classes prefer PR-AUC, recall at fixed precision or expected cost over accuracy.
  • What is data leakage? Information unavailable at prediction time used in training — target encoding without folds, future data in time series, duplicates across splits.
  • Why does gradient boosting work well on tabular data? Sequential trees fit residuals, handle mixed feature types and non-linearities with little preprocessing.
  • How does attention work? Each token builds a weighted average of value vectors, with weights from query–key similarity scaled by the square root of the dimension and normalised by softmax.
  • What does calibration mean? Predicted probabilities match observed frequencies; check with reliability curves, fix with Platt scaling or isotonic regression.

Coding round

Expect standard algorithm questions at an easier level than pure software roles, plus practical tasks: implement k-means or logistic regression with NumPy, compute a metric from scratch, or transform a messy dataset with pandas. Write tests for edge cases out loud.

ML system design: a walkthrough

Use the same skeleton for any prompt, for example "Design fraud detection for card payments":

  1. Clarify the goal, constraints and the cost of false positives versus false negatives.
  2. Define labels and how they arrive (chargebacks come weeks later).
  3. Features: transaction, user history, device, aggregates over time windows — computed without leakage.
  4. Model: start with gradient boosting, calibrated scores and a threshold chosen by cost.
  5. Serving: latency budget, online feature store, fallbacks when features are missing.
  6. Evaluation: offline metrics by time-based split, then A/B or shadow mode.
  7. Monitoring: data drift, score distribution, delayed label quality and retraining policy.

Project deep dive

  • Explain why you chose the metric and the validation scheme.
  • Show an error analysis and what you would try next.
  • Be honest about what did not work.

A six-week preparation plan

  1. Week 1: classic ML theory and metrics; write one-paragraph answers to common questions.
  2. Week 2: coding practice daily; implement two algorithms from scratch.
  3. Week 3: deep learning and transformers; re-derive backprop for a small network.
  4. Week 4: ML system design — three full walkthroughs on paper.
  5. Week 5: your projects — a short write-up and a five-minute story for each.
  6. Week 6: mock interviews and fixing weak spots.

Practising with feedback

Quantoros has an interview trainer with 2,000 questions across ML, statistics, Python, SQL and quant topics, with spaced repetition that brings back the questions you missed. Mock interviews with a mentor, in English or Russian, simulate the real rounds and end with specific feedback.

Resources

Questions

How long should I prepare for ML interviews?

Four to eight weeks is common if you already work with ML; longer if you are switching fields.

Is LeetCode necessary for ML roles?

Some practice is, but ML roles usually ask easier algorithm questions and more practical data and modelling tasks.

How do I prepare for ML system design?

Learn one reusable structure and practise it on several prompts: recommendations, search, fraud, forecasting.

Can I do mock interviews in English?

Yes, Quantoros runs mock interviews in English or Russian.

Mentoring directions