Activation was stuck at 45% and recommendations were a vendor black box nobody could see, tune, or own. Replaced it with a modular, explainable in-house platform spanning ranking, presentation, and a GenAI discovery assistant.
Role Built and owned the in-house recommendation platform end-to-end — architecture, ranking, experimentation, and the GenAI assistant
Recommendations are a core activation lever, but the vendor running them gave product and marketing no way to see, test, or improve them.
Only 45% of users reached the Activation threshold of two minutes of viewing, and recommendation lanes weren't moving that number.
A third-party recommender couldn't be inspected, tuned, or tied to Joyn's own catalog and editorial signals.
With nobody able to see or test ranking logic, engineering, data science, product, and marketing had no shared roadmap.
Cost, opacity, and a monolithic design meant every fix required a vendor ticket instead of an internal experiment.
The vendor contract carried a fixed annual cost regardless of whether the recommendations actually performed.
Experimentation ran through Optimizely, separate from ranking, so testing new recommendation ideas was slow and disconnected.
The existing setup wasn't designed to extend cleanly to new markets like Switzerland and Austria.
The recommendation engine was rebuilt as self-contained layers so teams could test and ship independently, then extended with automated personalization and a GenAI assistant.
Split the recommender into candidate generation, ranking & personalization, and presentation, letting each layer be built and tested in parallel.In-house recsys · modular architecture
Reordered existing lanes by recency and user signals, and shipped a dedicated machine-curated discovery lane for new users.Presentation layer · cold-start
Built A/B testing directly into the platform API to replace Optimizely, and shipped a Streamlit tool exposing ranking scores with editorial override.Streamlit · A/B framework
Detected recurring viewing habits to prioritize matching content at predicted times, and launched a conversational discovery assistant with trace-based evaluation.GenAI · pattern detection
The same modular architecture rolled out to Switzerland and Austria on shared infrastructure, and the embedded A/B framework now serves non-recommendation use cases across the product.
A machine-curated discovery lane for new users raised activation by 10 percentage points from a sub-25% baseline.
A new model identifies which users are most likely to watch specific content, powering lookalike and retargeting campaigns for CRM.
A conversational assistant for content discovery reached 100,000 users in four months, improving via trace-based evaluation.
Monitoring dashboards track recommendation quality and system health, enabling fast rollback of underperforming models.
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 →