Recommendation Engines

Their existing logic was mostly “most popular” and “same genre”; watch time and return visits were flat. We built a recommendation pipeline: implicit and explicit signals (views, completes, likes, skips), a ranking model that balanced relevance and diversity, and A/B tests on placement and algorithm. We also added “because you watched X” and “trending in your region” modules and tuned for long-term engagement, not just click-through.

Over two quarters, average session length and 7-day return rate both improved; the new “for you” surface became the top entry point for returning users. We handed over the pipeline and documentation so their team could iterate on features and models.