MediaStream

MediaStream's recommendation engine relied solely on viewing history and genre tags, missing the rich visual and thematic signals embedded in content itself. Subscribers reported that recommendations felt repetitive, and the platform's 30-day retention rate had stalled at 61%. The content team estimated that 40% of the catalog was virtually undiscoverable through existing metadata alone.

Mixpeek processes every title's trailer, thumbnail, synopsis, and subtitle track to generate rich multimodal embeddings. A multi-stage retriever pipeline combines visual similarity, narrative theme matching, and collaborative filtering signals to surface unexpected-yet-relevant recommendations. The system also powers a natural language search bar where users describe what they want to watch in plain English.

The rollout started with the documentary catalog, where sparse metadata was the biggest problem. Batch ingestion of 28,000 titles completed in 72 hours using a Ray Data pipeline. The retriever was integrated into the existing recommendation API via a lightweight adapter layer, requiring zero changes to the front-end clients. A/B testing ran for eight weeks before full deployment.