2020
Showroom Livestream Recommendation Engine
Replaced Showroom's naive popularity-based recommendations with an AI-driven system using content and user embeddings to match viewers with relevant livestreamers based on actual viewing behavior.
The Challenge
Showroom, a major Japanese livestreaming platform, recommended streams based solely on viewer count — a popularity metric that reinforced a winner-take-all dynamic, burying smaller creators and showing every user the same content regardless of their interests. The platform needed personalized discovery, but the team had no ML expertise and the solution needed to run cost-effectively at scale.
The Approach
Self-taught PyTorch, ONNX, and FastAI to build the system from scratch. Built content and user embeddings from viewing history data, replacing the crude viewer-count heuristic with a model that learns what each user actually watches and surfaces relevant streamers they haven't discovered yet. Designed a serverless deployment on AWS Lambda to keep infrastructure costs low. Mentored the engineering team on ML fundamentals to ensure the system could be maintained and iterated on after handoff.
The Results
Shipped a production recommendation engine that moved Showroom from one-size-fits-all popularity rankings to personalized streamer discovery. The system runs serverlessly on AWS Lambda, keeping operational costs minimal. Established the team's first in-house ML capability through hands-on mentoring, enabling continued iteration without external dependency.