ML product recommendation engine for an online store (API + offline evaluation)

We run a mid-sized e-commerce store with tens of thousands of active SKUs and a steadily growing catalog, and we need a machine-learning recommendation engine to replace the naive category-based blocks we use today. The two placements we care about most are a "related products" block on each product page and a "you may also like" block on the cart and post-purchase pages. Recommendations must be driven by real user behavior and purchase history: views, add-to-cart events, wishlists, completed orders and returns. We already log these events, so the first part of the job is designing a clean pipeline that ingests our event stream and order database into a training-ready dataset, with sensible handling of anonymous versus logged-in users and of cold-start items that have little or no interaction data yet.

The engine itself should be served through a documented API that our front-end and back-end can call in real time. For a given product or user session it must return a ranked list of item identifiers with scores, respond within a tight latency budget, and gracefully fall back to popularity-based or category-based results when the model has no confident prediction. We expect you to propose and justify the modeling approach — collaborative filtering, content-based signals, or a hybrid — but the choice must be explainable and reproducible, not a black box we cannot maintain. Business rules matter too: the API needs to exclude out-of-stock items, avoid recommending something the user just bought, and support simple filters such as price band and category.

Just as important as the model is honest offline evaluation. Before anything ships we want to measure recommendation quality on historical data using a proper train/test split by time, reporting standard ranking metrics and a plain-language summary of what improved over the current baseline. Deliverables include the trained model and inference service, the data pipeline, the evaluation report and reproducible scripts, and clear documentation for retraining as new data arrives.

— Data pipeline turning behavioral events and order history into training data
— Recommendation model (collaborative, content-based or hybrid) with cold-start fallback
— Real-time API returning ranked, scored recommendations for product and session
— Business rules: stock, price, category filters, exclude just-purchased items
— Offline evaluation with time-based split and ranking metrics vs. baseline
— Documentation and reproducible retraining scripts