Machine Learning Engineer interview
at Stripe.

Tell-me-about-a-time stories on ownership, conflict, and impact — scored against this company's real values.

1. 01 Tell me about a time a model you owned caused real user harm, such as blocking legitimate transactions, and how you responded.
2. 02 Describe a time you pushed back on a flashy modeling approach in favor of a simpler, more rigorous solution.
3. 03 Walk me through a time you took ownership of an ML system in production end-to-end, from data to monitoring.
4. 04 Tell me about a time you had to make a fraud or risk tradeoff that balanced a user's experience against financial loss.

The hard screen for the craft itself — talked through out loud, not whiteboarded in your head.

1. 01 A fraud model's recall drops sharply after a data refresh with no code changes and a flattened precision-recall curve; diagnose the likely cause and a fix.
2. 02 Given a stream of card transactions, walk me through the features you'd engineer to detect fraud in real time and why each is predictive.
3. 03 How would you choose and justify the decision threshold for a fraud model given the asymmetric cost of a false block versus a missed fraud?
4. 04 How would you handle severe class imbalance and label delay (chargebacks arriving weeks later) when training a fraud classifier?

Open-ended design of the systems this company actually runs, with the interviewer probing your tradeoffs.

1. 01 Design a real-time fraud-detection system that scores card payments within the authorization latency budget.
2. 02 Design the feature store and serving pipeline that guarantees training-serving consistency for fraud models.
3. 03 Design a system to monitor a deployed fraud model for drift, feedback loops, and degraded precision over time.
4. 04 Design a retraining and safe-rollout pipeline for fraud models that prevents a bad model from blocking legitimate payments at scale.