# ML System Design: Movie Recommendation Model and Pipeline (AI-Assisted Round) This is a hands-on, AI-assisted coding round: you are allowed to use AI coding tools (e.g., an AI pair-programmer in your editor) while you build. Design and sketch the implementation of a **movie recommendation model and its data/training/serving pipeline** for a streaming service. You should both reason about the system design and produce working scaffolding code for the core pieces — the data preparation, the model, the training loop, and the serving/inference interface — explaining your choices as you go. Concretely, build toward a system that, given a user, returns a ranked list of movies they are likely to watch. Cover the full pipeline: ingesting interaction data, generating features/labels, training a recommendation model, evaluating it, and serving recommendations at low latency, including how you handle retraining and cold start. ### Constraints & Assumptions - Catalog on the order of $10^5$ movies; tens of millions of users. - Primary interaction signal is watch events (plus optional explicit ratings, search, browse). - Online recommendations must return within a low-latency budget (tens of milliseconds) for a few hundred items. - The model is retrained on a regular cadence (e.g., daily/weekly) on fresh interaction logs. - New users and new movies appear continuously (cold start). - Because AI coding tools are permitted, the interviewer expects you to move quickly to runnable scaffolding and to critically review the AI-generated code rather than accept it blindly. ### Clarifying Questions to Ask - What surface is this for — the personalized home feed, "because you watched X," or post-play "up next"? It changes the candidate set and latency budget. - What is the optimization objective — predicted watch probability, expected watch time, or a multi-objective blend with diversity/freshness? - What interaction signals are available, and are there explicit ratings or only implicit watch/skip events? - What is the online latency and QPS budget, and is there an existing feature store / serving infra to reuse? - How fresh must recommendations be — do we need near-real-time updates from a session, or is daily/weekly retraining acceptable? - Since this is AI-assisted, are you evaluating the final design, my code-review judgment over AI output, or both? ### What a Strong Answer Covers ```premium-lock What a Strong Answer Covers ``` ### Follow-up Questions - Walk through your two-stage design's latency budget end to end: where does the time go in retrieval vs ranking, and what do you precompute offline vs compute per request? - For implicit feedback, how exactly do you construct negatives, and what goes wrong if you sample them naively (e.g., uniformly over the whole catalog)? - You used an AI tool to generate the training loop. Show me a specific bug or train/serve-skew risk you would specifically check the generated code for, and how. - How do you prevent a feedback loop where the recommender keeps recommending what it already recommends, starving exploration and new content? - A movie is added to the catalog today with zero interactions. Trace exactly how it can still appear in someone's recommendations within minutes.