Product Recommendation Engine
A recommendation system for suggesting relevant products to users.
The problem
In a large catalog, users cannot find relevant products without personalized suggestions.
My role
Prototype — design & build
Product Recommendation Engine
A recommendation system for a large product catalog. It surfaces relevant products per user and explores the trade-off between personalisation quality and the cold-start problem for new users and new products. It was built alongside the evaluation SDK, because a recommender you can't measure is a recommender you can't improve.
Key decisions
Build the measurement alongside the model, not after it. Every variant was compared with the same repeatable evaluation instead of judged by eye, because "it looks reasonable" is not a metric.
Evaluate offline first. Variants were scored on test data before anything faced a user, which is slower to celebrate and much cheaper to be wrong with.
Outcome
Working prototype that produced personalised suggestions over a test catalog and could be compared version to version; deliberately kept at prototype stage.
Limitations
Prototype-stage: evaluated offline on test data, not against live user behaviour.
Stack
Next
A production version would need online evaluation, cold-start handling and a feedback loop from real interactions.