AI HR
Prova

Interview questions

Data scientist interview questions

Data science interviews test whether your work changes decisions, not just whether you know the algorithms. Expect to defend a model choice, design an experiment on the spot and explain a result to someone who will never read your notebook.

Rehearse this interview by voice2 minutes without an account, 3 more after sign-up. No card.

Questions by stage and what a strong answer includes

  1. 01

    Is your background closer to product analytics or machine learning?

    Strong answer: An honest split with examples of each, so the recruiter can place you in the right loop. Mention tools and the scale of data you worked with.

  2. 02

    Which project best shows your impact?

    Strong answer: One project, the business question, the method in one sentence and the decision or metric it changed.

  1. 03

    Tell me about a time your analysis was ignored or challenged.

    Strong answer: How you checked your work, reframed it in terms of the stakeholder's goals, and what happened. Influence, not just being right.

  2. 04

    Describe a model that performed well offline but failed in production.

    Strong answer: The cause, such as leakage, drift or a wrong metric, how you noticed, and what you changed in validation or monitoring.

  1. 05

    How would you design an A/B test for a new onboarding flow?

    Strong answer: Primary metric and guardrails, unit of randomization, sample size and duration, checks for novelty effects and sample ratio mismatch, and a decision rule set in advance.

  2. 06

    Why did you choose that model over a simpler baseline?

    Strong answer: Comparison against a baseline with the metric that matters to the business, plus costs: latency, interpretability, maintenance. Willing to choose the simpler model.

  3. 07

    Our weekly active users dropped 5%. How do you investigate?

    Strong answer: Check data and tracking first, then segment by platform, region, cohort and acquisition channel, look for releases or external events, and state hypotheses before digging.

  4. 08

    How do you explain precision and recall to a product manager?

    Strong answer: A plain example tied to their product, like fraud flags or recommendations, and what each type of error costs the business.

  5. 09

    How do you take a model from notebook to production?

    Strong answer: Reproducible pipeline, versioning of data and models, tests, deployment approach, and monitoring for drift and performance, plus who owns retraining.

  1. 10

    What data problem would you tackle first here?

    Strong answer: A thoughtful hypothesis based on what you know of the product, and the questions you would ask before committing.

  1. 11

    What compensation range are you expecting?

    Strong answer: A researched total compensation range by level and location, and openness on base versus equity if the total fits.

One answer, before and after

Describe a model that performed well offline but failed in production.

Weak answer

We built a churn model with really high accuracy, but in production it didn't work as well. Data in the real world is messy, so we retrained it.

Strong answer in STAR

Situation
At a subscription company I built a churn model that scored 0.86 AUC on the holdout set. After launch, the retention team said the customers it flagged were mostly already gone.
Task
I needed to find out why the model was useless in practice and get a version the retention team could act on.
Action
I traced the features and found that 'days since last login' was computed at the time of the data pull, not at prediction time, which leaked the outcome. I rebuilt the training set with point-in-time features, set the prediction window to 30 days ahead, and agreed with the retention team that the useful metric was precision in the top 500 scored customers each week.
Result
The new model scored a lower 0.74 AUC, but precision in the top 500 rose to about 40%, and a four-week test with retention offers reduced churn in the targeted group by about 3 points versus control. Point-in-time feature checks became part of our model review template.

Why it works

The weak answer blames messy data. The strong one names the exact leakage, redefines success in business terms and proves value with an experiment. The numbers are illustrative: use your own.

How a practice session works

  1. 01

    Pick a stage, paste the job posting

    Recruiter screen, HR interview, hiring manager, final round or offer talk. Add the posting and your resume so the questions are about your role, not generic.

  2. 02

    Talk it through by voice

    The interviewer speaks the language you chose, interrupts, follows up and digs into details like on a real call. If the connection drops, you pick up where you stopped.

  3. 03

    Get the review and drills

    A score, 7 skills, what you said next to a stronger way to say it, plus pace, filler words and talk share. Weak spots turn into short drills.

Rehearse this interview by voice

Related roles