RL Trading Platform

An algorithmic trading platform that traded real money through Alpaca's API for several months in 2025 and now runs as my backtesting and research system. The hot path is Go, ingesting over 100,000 market messages per second per core. While it traded live, the full round trip, from data ingress through ML inference to a sent and confirmed order, completed in under 100 milliseconds, and most of that was network travel time to the broker's API. The strategy layer is Python and PyTorch, where reinforcement-learning models are trained offline and promoted into production behind a versioned interface, so a bad model can be rolled back without touching the execution engine. Redis carries hot state between services, everything ships in Docker, and the stack ran on AWS EC2 with 99%+ uptime through the live period. This is the project where the embedded habits pay off in cloud software: tight latency budgets, defensive failure handling, and the assumption that anything that can fault eventually will. The repository is private because it holds my strategies; I am happy to walk through the architecture in conversation.