PitWall-AI
Agentic F1 race-strategy engine — five cooperating agents (LangGraph) making per-lap pit decisions with PyTorch tire models, genetic algorithms, and an LLM race engineer.
Overview
An end-to-end Formula 1 race-strategy simulator that makes per-lap pit-stop and compound decisions with the reasoning of a real pitwall — fully automated. Replay any 2023–2025 race at up to 1000× speed, or point it at a live Grand Prix.
Problem
Race strategy is a multi-agent reasoning problem under uncertainty: live telemetry, rival behavior prediction, combinatorial strategy search, and clear communication — all per lap, in real time.
Approach
Scout ingests live telemetry from the OpenF1 API; Spy predicts rival pit stops with a Bayesian hazard model; Strategist evolves the optimal remaining-race plan with a DEAP genetic algorithm over a PyTorch tire-degradation network; Ghost Car scores strategies against the real driver's decisions; Principal writes team-radio briefings via LLaMA-3.3-70B. Everything streams to InfluxDB 3 Core (SQL-native, Apache Arrow) and renders on a 32-panel Grafana dashboard.
Impact
A flagship agentic-AI artifact: real multi-agent orchestration, ML + evolutionary search, time-series engineering, and LLM integration in one system — presented at an InfluxData webinar (July 2026).