Design a Recommendation System (Two-Tower, Candidate Gen, Ranking)

How an e-commerce platform narrows 1 crore products down to 20 personalised items in 47 ms using two-tower vector embeddings, HNSW graph search, dual-tier feature stores, and GPU ranking under a strict latency SLA.

Architectural blueprint for an enterprise recommendation engine serving 5,000 requests a second at p99 under 50 ms. Covers multi-stage funnel candidate retrieval, asymmetric two-tower embeddings, sub-12ms HNSW ANN indexing, online streaming feature stores in Redis, Triton GPU multi-task neural ranking, and MMR diversity re-ranking.