Each question below includes what the interviewer is actually testing. That matters more than the answer itself — once you know what is being probed, you can answer from your own experience instead of reciting someone else’s script, which is obvious to everyone in the room.
Whether you have crossed the deployment gap at all. The answer separates this role from data science more reliably than any other question.
Spend most of the answer past the training step: how it was served, how you validated it on live traffic, what monitoring you added, and what broke afterwards. If you stop at "we got 0.91 AUC", the interviewer concludes the model never shipped.
The most diagnostic question in this loop. It requires understanding train/serve skew, feedback loops, and metric mismatch simultaneously.
Work through the candidates: training/serving skew from features computed differently in each path, distribution shift between your offline sample and live traffic, an offline metric that does not correspond to the business metric, or an experiment underpowered for the effect size. Say which you would check first and why.
Seniority. Junior engineers apply ML because it is the tool they have; senior engineers price it against the alternative.
Ask whether a heuristic gets most of the value at a fraction of the cost, whether you have labels and enough of them, whether the cost of a wrong prediction is tolerable, and who maintains it in a year. The strongest answer includes a time you argued against a model and shipped rules instead.
Whether you have owned a model after launch. Models decay silently — nothing errors.
Distinguish the layers: input drift, prediction drift, and true performance measured against delayed labels. Explain what you monitor when labels arrive late or never, mention alert thresholds you actually set, and describe your rollback path.
Explainability and communication, and increasingly a compliance requirement rather than a nice-to-have.
Give the feature-level explanation without jargon, be honest about what the model cannot tell you (correlation, not causation), and describe the appeal or human-review path. In regulated contexts, mention what you keep on record to defend the decision later.
Fundamentals, and it catches the accuracy mistake before it reaches the resume.
Stop using accuracy; move to precision/recall, PR-AUC, or a cost-weighted metric tied to the real consequence. Then address the imbalance — class weights, resampling, threshold tuning on a validation set — and pick the operating point from business cost rather than from the F1 peak.
None of this matters if your resume is filtered out first. The most common cause is not a weak background — it is a document that screening software cannot read correctly, or one that lists duties where it should show results.
Check your resume against machine learning engineer keywords or run a full scan. Both are free and take about three seconds.