Mock interviews by role · Data, AI & Cloud
Data Scientist mock interview
Data scientist interviews separate people who fit a model from people who can say when the model should not be used. Leakage, the decision the model serves, and a simple baseline belong in almost every answer.
If you have not shipped a model, use a project and say so. Describe the baseline you beat and the way the result could be wrong.
What the loop usually looks like
Screen
30 min
A modeling problem and the decision it supported.
Technical
45–60 min
Experiment design, metrics, or a take-home review.
Applied judgment
45 min
Failure modes, fairness, or a stakeholder who wants certainty.
What interviewers score
- Decision framing
- Baselines
- Leakage and validity
- Limits of the result
10 data scientist interview questions and model answers
Technical
1. How do you know a model is leaking future information?
What they are testing: Validity.
I check whether any feature is only known after the outcome, including aggregates computed on the full timeline. I split by time, not at random, when the product will score future events. A random split that looks brilliant is a warning, not a win.
Technical
2. What baseline do you beat before you trust a complex model?
What they are testing: Humility.
A majority-class rule, a single strong feature, or the current heuristic the business already uses. If the complex model barely beats that, I ship the simple rule and explain why.
Domain
3. Design an experiment for a new ranking of search results.
What they are testing: Experiment design.
I assign users, not queries, so one person does not see both ranks. The primary metric is the successful task, not just clicks. I decide the sample and the stop rule before launch, and I watch a guardrail metric such as refunds or long-latency.
Situational
4. The model is accurate and still harms a segment. What do you recommend?
What they are testing: Product judgment.
I show the error rate by segment next to the average. If one group is far worse, I do not hide it in the mean. Options are a threshold change, a human review for that group, or not shipping. Accuracy alone is not the decision.
Behavioral
5. Tell me about a model you did not ship.
What they are testing: Judgment.
It beat the baseline offline and failed on a week of live traffic because the features were delayed. We kept the heuristic. The useful output was the monitoring that showed the feature lag, which we did not have before.
Technical
6. How do you explain a model to a stakeholder who asks for the why?
What they are testing: Communication.
I describe the decision, the inputs a human would also use, and one example where the model and the human disagree. I do not start with the algorithm name.
Domain
7. When is a metric the wrong objective to optimize?
What they are testing: Goodhart.
When people can raise it without improving the outcome, such as clicks that are mis-taps, or handle time that falls because agents hang up. I pair it with a quality check the team cannot game as easily.
Behavioral
8. Describe a disagreement with an engineer about training data.
What they are testing: Collaboration.
They wanted every historical row. I excluded a period when the logging was broken because those labels were noise. We documented the exclusion in the dataset note so the next train did not silently add it back.
Technical
9. How do you monitor a model after launch?
What they are testing: Operations.
I track the score distribution, the input drift on a few key features, and the business metric the model was meant to move. A stable accuracy report with a moving input is not enough.
Situational
10. You have a small labeled set. What do you refuse to claim?
What they are testing: Scope of evidence.
I refuse a precise accuracy number as if it were stable. I say the interval is wide, I show the baseline, and I ask for the cheapest way to label more of the cases we would actually decide on.
Frequently asked questions
Do I need a research background?
No. Many roles want someone who can frame a decision and ship a simple model with monitoring. Lead with the decision and the baseline.
How much math should I derive on a whiteboard?
Know the metrics you cite well enough to define them. Deriving a loss from scratch is less common than explaining an experiment or a leak.
What if my best project is coursework?
Use it, label it as coursework, and spend your time on the baseline and the mistake in the first split. That is the part interviewers trust.
Can follow-up questions probe the experiment?
Yes. In a data scientist mock, expect a follow-up on the unit of randomization or the guardrail metric if you skipped them.