Design a scalable AI recommendation system.

### Signal to interviewer

I can design recommendation systems that scale while balancing relevance, diversity, and platform constraints.

### Clarify

I would clarify recommendation surface, user intent patterns, content safety requirements, and acceptable freshness delay.

### Approach

Use a candidate-ranking pipeline: generate candidates fast, rank with multi-objective scoring, then apply policy constraints and experimentation controls.

### Metrics & instrumentation

Primary metric: qualified engagement per session. Secondary metrics: retention lift, catalog coverage, and cold-start performance. Guardrails: harmful content exposure, creator fairness imbalance, and latency regressions.

### Tradeoffs

Tighter personalization improves near-term engagement but can reduce diversity. Aggressive exploration improves discovery but may lower immediate conversion.

### Risks & mitigations

Risk: feedback loops amplify narrow preferences; mitigate with exploration caps. Risk: stale recommendations; mitigate with incremental feature refresh. Risk: gaming by content producers; mitigate with integrity scoring.

### Example

In a short-video feed, candidate generation uses recent watch signals while ranking enforces diversity and policy-safe filtering before final delivery.

### 90-second version

Build a scalable two-stage recommender: fast candidate retrieval, smart ranking, and strict policy controls. Optimize for qualified engagement while protecting diversity and safety.