Top Machine Learning Engineer Interview Questions & Answers

Requirements clarification: • Serve personalized product recommendations in < 100ms • Handle 10K requests/second at peak • Incorporate user behavior, purchase history, and product metadata Architecture: 1. Offline pipeline (batch, runs daily) • Feature engineering from clickstream and purchase data • Train collaborative filtering + content-based models • Generate candidate sets per user segment • Store in a feature store (Feast/Tecton) 2. Online serving (real-time) • Candidate retrieval: pre-computed similar items + ANN (Approximate Nearest Neighbor) search via Faiss/Milvus • Ranking: lightweight ML model (gradient-boosted tree) scoring candidates using real-time features (current session, cart contents) • Result: top-K ranked items returned in < 50ms 3. Monitoring • Track CTR, conversion rate, and diversity metrics • Data drift detection on input features • A/B testing framework for model comparisons Key tradeoff: Offline models provide coverage; online ranking provides personalization. The two-stage architecture (retrieve then rank) is the industry standard.

Types of drift: • Data drift (covariate shift) — Input data distribution changes. E.g., your model was trained on US users but now serves international users with different behavior patterns. • Concept drift — The relationship between inputs and outputs changes. E.g., what constitutes a "good" recommendation changes due to seasonal trends. • Prediction drift — Model output distribution changes, even if individual predictions may still be correct. Detection methods: 1. Statistical tests on input features (KS test, PSI — Population Stability Index) 2. Monitor prediction distribution over time 3. Track business metrics (CTR, conversion) and set thresholds 4. Feature attribution monitoring (SHAP value distributions) Handling drift: • Automated retraining — Scheduled (daily/weekly) or triggered by drift alerts • Online learning — Model updates incrementally with new data • Fallback models — Switch to a simpler, more robust model during high-drift periods • Human-in-the-loop — Alert data scientists when automated systems can't resolve drift