horse-racing-ai-jev

Horse Racing AI is a horse racing simulator that integrates real-time AI prediction using the TypeSafe Jev model. The project combines a React frontend with a Python-based AI agent to predict race outcomes before and during a race. The Jev model is the only prediction model used, and it receives race state information and a single typed question to return win probabilities per horse. The system is designed to predict race changes in under a second.

The frontend generates a race and calls the predict_race tool with a public briefing that includes weather, distance, horses, and rules. Jev returns pre-race win probabilities. During the race, the frontend periodically builds a live snapshot with distance run, observed pace, finish places, falls, and visible incidents, and calls predict_race again. The JavaScript engine determines the winner, and the UI compares Jev's pre-race and live predictions using metrics like favorite, probability on the winner, and Brier score. Jev never sees engine internals such as formulas, energy, form, or results.

The project includes an MCP agent built with Python, FastMCP, Pydantic, and uvicorn, and it exposes tools like predict_race and predict_races. The agent sends structured data to Jev, including rules, race details, horses, and live standings with computed gaps and positions. The system is run via Docker Compose and requires a TypeSafe API key for operation. Tests are provided and do not call TypeSafe, using a fake API instead.