In this video, I break down Netflix-style recommendation system design with the classic two-stage setup: a retriever and a ranker. The big idea is simple—at Netflix scale you can’t score every video for every user in real time, so you first retrieve a small candidate set fast (from millions to hundreds/thousands), and then you run a heavier ranking model to order those candidates for the final homepage rows. I walk through how to think about user signals, item features, and why “retrieval vs ranking” is not just a modeling choice—it’s a latency + cost + quality trade-off. I also connect the dots to practical engineering: what features typically go into each stage, how embeddings help with candidate generation, and why ranking models often look like CTR/watch-time predictors with richer context (device, time, session intent, freshness, diversity). If you’re building your own recommender (or even a RAG-style retrieval pipeline), this is the same system-design thinking: narrow the search space first, then spend compute where it matters. If you want to implement this end-to-end, I’ve shared my code and updates on GitHub and I publish deeper breakdowns on Medium—those links are in the description.