Uber Machine Learning Engineer Mock Interview

Tell us about a time you took a machine learning model from offline experimentation to production at scale—what were the biggest technical and product risks (latency, reliability, data drift, feedback loops), and how did you validate that your solution worked for real Uber use cases?

Uber relies on ranking and decisioning systems for matching riders and drivers. Walk through how you would design, train, and evaluate a model for improving marketplace matching (e.g., ETA or acceptance likelihood), including how you’d choose metrics that reflect the business objective rather than just offline accuracy.

Describe a situation where your model’s performance degraded after deployment. What signals did you monitor, how did you diagnose whether it was due to data shift, labeling changes, instrumentation bugs, or non-stationarity, and what mitigation plan did you implement?

Imagine you’re asked to reduce tail latency for a real-time inference service that powers Uber’s recommendations or dispatching. What engineering approach would you take to meet latency SLOs (model compression, feature store optimization, caching, batching, quantization, or approximate methods), and how would you measure the trade-offs?

A stakeholder wants to deploy a new ML approach for a specific Uber market, but the available historical data is limited and biased (cold-start and differing user behavior). How would you handle limited data and bias—what would you try first, and how would you structure an experiment or rollout to ensure we don’t harm rider/driver outcomes?