Personalised Health Recommendations for 2M+ Users

Problem

Engagement with health content was low because recommendations weren't grounded in either the user's clinical context or actual behavioural signals. We needed a recommender that respects both.

Approach

1. 1.Engineered behavioural features alongside clinical attributes; trained ranking models with offline-online consistency checks.
2. 2.Layered policies on top of the ranker to respect clinical safety constraints (e.g. surface vs suppress conditions).
3. 3.Closed the loop with A/B tests and engagement metrics tied to long-term outcomes, not just CTR.