AI system design interviews require a structured approach. Use the PASSTR framework: Problem definition, Architecture overview, Scale estimation, Storage design, Training pipeline, Real-time serving. For recommendation systems: start with candidate generation (collaborative filtering + content-based), then ranking (two-tower neural network), then post-processing (diversity, business rules). For fraud detection: use a two-stage approach with fast rules (<1ms) followed by GBM scoring (<10ms) and optional deep learning for high-value transactions. Scale estimates matter: 500M MAU means ~5800 recommendation QPS. Interview tips: ask clarifying questions about business metrics, discuss tradeoffs explicitly, address cold start and distribution shift, always mention monitoring.