The existing storefront showed the same "popular items" to everyone. Personalisation products on the market either required sending full user data to a third party, or introduced latency that hurt page-load metrics.
The team needed a recommender they controlled, fast enough to render inline, and good enough to lift conversion without spamming users with items they had already bought.
A hybrid recommender combining content-based features (product attributes, images, embeddings) with collaborative signals (co-views, co-purchases, session sequences).
Models are retrained on a daily cadence, served behind a thin Go API at single-digit millisecond latency, with feature flags so the team can A/B test ranking strategies.
A feature pipeline aggregates product attributes and behavioural signals (views, add-to-carts, purchases) into an offline feature store.
Candidate generation combines ANN search over product embeddings with collaborative-filtering signals; a lightweight ranking model scores candidates per request.
A Go service serves recommendations behind a CDN, with feature flags and online metrics for ranking experiments.
