Build a Two-Tower Retrieval Engine

• ✓Build a two-tower retrieval model — separate user and item encoders trained into one shared embedding space
• ✓Train with in-batch negatives and sampled softmax, and correct the popularity bias they introduce
• ✓Measure retrieval honestly with Recall@K, NDCG@K, and catalog coverage against a popularity baseline
• ✓Index millions of item vectors with IVF and HNSW, and read the recall-versus-latency trade-off from your own measurements
• ✓Retrieve items that have never been interacted with, by putting content features in the item tower
• ✓Explain why production recommenders split retrieval from ranking, and measure what each stage contributes
• ✓Serve the engine behind an API and report the p99 latency it sustains under concurrency

Build the candidate-generation layer that sits under every large recommender — the stage that turns a catalog of millions into a shortlist of hundreds, fast enough to run on every request.