Machine Learning Engineer interview
at Uber.

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 you owned an ML model end-to-end from offline experimentation through production deployment and on-call monitoring.
2. 02 Describe a time you made a bold modeling bet that improved a core metric but required convincing skeptical stakeholders to trust it.
3. 03 Give an example of when a deployed model degraded in production and how you took ownership of diagnosing and fixing it.
4. 04 Tell me about a time you had to balance model accuracy against latency and cost constraints for a real-time, customer-facing system.

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

1. 01 How would you monitor, evaluate, and validate the real-time performance of a newly deployed ETA prediction model in production?
2. 02 How would you determine whether one million historical trips is enough data to train an accurate ETA model across routes, times, and weather?
3. 03 How would you frame and engineer features for a model that ranks which driver to dispatch to a given ride request?
4. 04 How would you detect and mitigate training-serving skew for a model whose inputs are streamed live driver and traffic features?

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

1. 01 Design a real-time ETA prediction system that serves low-latency predictions across every active trip in a city.
2. 02 Design an ML platform like Michelangelo that supports feature storage, training, deployment, and monitoring for many models across Uber.
3. 03 Design a fraud detection system that scores rides and payments in real time without adding noticeable latency to the trip flow.
4. 04 Design the serving infrastructure for a driver-rider matching model that must rank candidates within tight latency budgets at peak load.