The Challenge
What Discover Was Facing
Discover needed a recommendation engine that could surface genuinely relevant results for London users across a catalogue with specific local context — neighbourhood relevance, cultural nuance, and real-time availability — that generic recommendation APIs from Google and AWS Personalize were fundamentally unable to model. Third-party APIs produced recommendations that were statistically plausible but contextually wrong for a London audience. Tuning generic algorithms to London market specifics required ongoing engineering effort that negated any cost saving from using an off-the-shelf service.
The Solution
What We Built
We built a bespoke recommendation engine trained on London-specific engagement data, with a feature set that incorporated geographic proximity, neighbourhood context, time-of-day relevance, and inventory availability as first-class signals. The engine was designed to be retrained incrementally as new engagement data accumulated, without requiring a full retraining cycle. A/B testing infrastructure was built into the platform from day one, allowing the product team to evaluate algorithm changes against live traffic without engineering involvement.
Results