Search is Joyn's second-largest acquisition source, with 60-70% of searchers converting onward — but Algolia ranked on static catalog text alone, and dead-end queries had no recovery. Rebuilt in two phases into a behavior-driven, self-improving ranking and fallback system, with zero added latency at millions-of-DAU scale.
Role Architected and led the search relevance rebuild — signal pipeline, GDPR governance, and hybrid reranking
High traffic through search wasn't translating into proportional relevance — vendor ranking couldn't see what users actually did next.
Search is Joyn's second-largest home page acquisition source, behind only the hero lane. Most sessions that reach it convert onward.
Results were ordered by static catalog text, with zero visibility into clicks, views, or watch-time. Relevance couldn't improve itself.
A near-miss query returned nothing and no alternative path, pushing users straight to bounce.
Three constraints — data access, compliance, and architecture — had to be solved together before relevance could improve.
Click-throughs, video views, and watch-time lived in Joyn's systems but never reached Algolia's ranking model.
Any pipeline sending user behavior to a third-party search vendor required consent handling and Legal sign-off.
Algolia's core results had no layer to blend keyword relevance with Joyn's own recommendation models.
First-party behavioral data closed the feedback loop with Algolia; a custom layer then handled fallbacks and reranking on top.
Built a pipeline transmitting click-throughs, video views, and watch-time from consented users directly to the search index.Algolia · event pipeline
Partnered with Data Science and Legal to validate consent scope and data sharing before any signal went live.Legal · Data Science
Embedded the new signals into Algolia's ranking model, shifting it from static metadata matching to behavior-driven relevance.Algolia · ranking model
Built a custom reranking layer on Algolia's results using in-house ML, and shipped personalized fallbacks for zero-result queries.In-house recsys · reranking
First-party behavioral data now feeds ranking continuously, scaling to millions of daily active users with zero latency degradation — relevance improves itself with every session.
Zero-result recovery preserves user intent instead of ending the session at an impasse.
Keyword relevance and ML-driven engagement signals now combine, outperforming either alone.
The in-house reranking model built for search now generalizes to other discovery surfaces.
Early Product, Data Science, Engineering, and Legal alignment became the template for shipping compliant data features fast.
Let's talk
Tell me your biggest data or AI bottleneck and I'll send you how I'd approach it — the same way I worked this problem. No résumé dump, no pitch. Just a useful conversation.
Start the conversation →