Use this guide to interview a Machine Learning Engineer — the role that turns models into reliable production systems, not just notebooks. Pick eight to twelve questions per session, keep the set consistent across candidates, and weight the production-systems group heavily: it separates ML engineers from pure researchers. Anchor questions in {{Your product or domain}} where possible.

ML fundamentals

• Explain the bias-variance trade-off using an example from a project you actually shipped.
• When would you choose a gradient-boosted tree over a neural network? What evidence would change your mind?
• How does regularization work, and how do you choose between L1, L2, and dropout in practice?
• What is data leakage? Describe a subtle leak you have found — or caused.
• How would you detect overfitting when the validation set is small?

Data and feature engineering

• Walk me through building a training set for {{A prediction problem in your domain}}. Where do labels come from, and what could poison them?
• How do you handle class imbalance beyond naive oversampling? When does it actually matter?
• Tell me about a feature pipeline you built. How did you keep training-time and serving-time features consistent?
• How do you version datasets and features so an experiment from six months ago is reproducible?
• What checks do you run before trusting a new data source?

Modeling and evaluation

• How do you choose an offline metric that predicts online impact? Tell me about a time the two disagreed.
• Design an evaluation for a model whose ground truth arrives weeks after the prediction.
• How do you set up an A/B test for a model change? What can go wrong besides insufficient sample size?
• Your model performs well on average but badly for one user segment. How do you find and fix that?
• When is a simple heuristic the right answer instead of a model? Give a real example.
• How do you communicate model uncertainty to a product team that wants a yes-or-no answer?

What good looks like: strong candidates tie every metric back to a product or business outcome, know exactly where offline metrics mislead, and volunteer at least one story where they killed their own model because a simpler approach won. Be wary of candidates who only discuss leaderboard-style accuracy.

ML systems and production

• Design the serving path for a model that must respond in under 100 milliseconds at high traffic. Where does feature computation live?
• How do you monitor a model in production? What is drift monitoring actually watching, and what do you do when it fires?
• Walk me through a training pipeline you owned: orchestration, retraining triggers, and rollback.
• Batch versus real-time inference: how do you decide, and what does each cost operationally?
• A model's production performance degraded but no code changed. What are your top hypotheses and how do you test each one?
• How do you roll out a new model version safely — shadow mode, canary, or something else — and why?

What good looks like: senior candidates treat the model as a small component inside a larger system. They bring up data contracts, monitoring, rollback plans, and failure modes unprompted, can quantify serving and retraining costs, and describe at least one production incident they diagnosed from symptoms back to a data cause.

Software engineering for ML

• How do you test ML code when behavior is partly statistical?
• How do you take notebook-quality research code to production quality without losing the researcher along the way?
• Tell me about a time you significantly cut training or inference cost. What did you change?
• How do you structure an ML repository so experiments do not rot into unmaintainable branches?

Collaboration and impact

• Tell me about an ML project that shipped and moved a business metric. What was your specific contribution?
• How do you scope a project with a product manager who is not sure whether ML is even needed?
• Describe explaining a wrong-looking model decision to a non-technical stakeholder.
• How do you mentor engineers new to ML without becoming the bottleneck for every modeling decision?