I built and shipped the recommendation service The Feed runs in production. It looks at what we know about a user: their Strava training profile, stated preferences, and past orders if they have them. From that it returns a ranked list of product categories and brands. It covers the whole user base: brand-new Strava-only users, which is the hard cold-start case, and returning customers where I can lean on their history. The storefront, onboarding, and email all call it over HTTP.

• Trained a LightGBM LambdaRank model over user × category and user × brand pairs, with fully decoupled training and serving that communicate only through a versioned model bundle
• Solved cold-start by ranking on Strava sport/geo and stated-preference features, then blending in recency-weighted purchase-history (RFM + category/brand affinity) once a customer has orders
• Deployed as a containerized FastAPI service on Railway with a mounted model volume, a weekly retrain cron, a backtest quality gate, and ~60s hot-reload of new bundles with zero downtime
• Achieved Hit@5 ≈ 0.88 on the production backtest, with per-request latency of ~40–80ms after warmup
• Fetches optional live purchase history from the ClickHouse CDP at request time, degrading gracefully to cold-start on any data-source error