Design a Cold-Start-Aware Recommender

You need to build recommendations for a new content app with sparse interactions and frequent new items. Propose a two-stage system (candidate generation + ranking) that is robust to user and item cold-start. Specify concrete model choices for each stage (e.g., matrix factorization or two-tower retrieval; gradient-boosted trees or DNN ranker), the exact loss functions and negative sampling strategy, key features (including user, item, and context), and how you will handle exploration vs. exploitation (e.g., Thompson sampling or epsilon-greedy with propensity caps). Define offline metrics and online KPIs, and design an A/B test that controls for selection bias using IPS/SNIPS or doubly robust estimators; write the estimator you would compute and describe guardrails for long-term outcomes. Finally, show how you will meet p50 < 50 ms and p99 < 200 ms end-to-end latency at inference, including feature store access patterns, caching, and approximate nearest neighbor retrieval.