Live Streaming Reinforcement Learning Recommendation System

Overview

A consumer-facing live streaming platform needed to improve real-time content recommendations under strict latency constraints. Traditional offline-trained models were slow to adapt to user behavior and changing content dynamics. The goal was to design a system that could learn continuously from live user interactions.

Achievements

A reinforcement learning–based recommendation system was deployed to production, adapting recommendations in near real time based on user feedback signals. The system improved engagement metrics while remaining stable under high request rates.

Responsibilities

• Designed the end-to-end recommendation architecture combining offline training and online learning.
• Defined reward signals based on user interactions (watch time, skips, engagement events).
• Built a real-time inference service with tight latency budgets.
• Implemented safeguards to prevent feedback loops and degraded user experience during exploration.
• Worked closely with product and backend teams to integrate the model into the live serving stack.