PITWALL
AI-powered pit stop strategy using real F1 telemetry, IBM Granite reasoning, Docling regulation parsing, and a Q-Learning agent trained on 23,400 real race decisions.
How PitWall Works
Three AI systems working in concert โ real telemetry data, regulation-aware language reasoning, and reinforcement learning trained on actual race outcomes.
Race Strategy Analysis
Select any season (2022โ2024), Grand Prix, and driver. FastF1 loads real telemetry. IBM Granite โ with Docling regulation context โ analyzes the complete race strategy.
- Real FastF1 telemetry โ not simulated data
- FIA regulation context injected via Docling
- 3-paragraph strategic breakdown from Granite
- Identifies optimal vs suboptimal decisions
- Suggests alternative strategies
Live Pit Window Advisor
Adjust sliders for current lap, tyre age, and lap time delta. IBM Granite gives a direct pit/stay-out recommendation simulating real pit wall decision support.
- Real-time Granite recommendation
- Tyre compound context aware
- Lap delta degradation analysis
- 2โ3 sentence direct reasoning
- Simulates actual pit wall workflow
RL Pit Optimizer
The Q-learning agent โ trained on 23,400 real F1 lap decisions โ outputs a pit recommendation with full Q-value transparency. Not a black box.
- Trained on 5 real F1 races (2023โ2024)
- 23,400 lap state โ decision pairs
- Full Q-value breakdown shown
- Confidence score from Q-value margin
- Explainable โ see exactly why it decides
Built with IBM AI
PitWall uses two core IBM technologies as required by the challenge โ Granite for reasoning and Docling for document intelligence.
Granite 4.1 8B
IBM's flagship open-source language model. Core reasoning engine for strategy analysis and pit window recommendations. Receives telemetry + regulation context, outputs explainable strategic analysis.
Docling
IBM's document intelligence library. Parses the FIA Sporting Regulations PDF, extracts tyre compound rules and pit stop regulations, injects this context into every Granite prompt.
FastF1 + RL Agent
Official F1 telemetry API for real race data (2022โ2024). Combined with a custom Q-Learning agent trained on 23,400 real lap decisions from 5 F1 races.
Try PitWall Live
The full application running live on HuggingFace Spaces. Select Monaco 2024 + LEC and click Analyze Strategy to see IBM Granite in action.
Live Race Simulation
Customize your strategy โ compound, pit lap, degradation rate. Watch the RL agent make lap-by-lap pit decisions. Pause anytime to analyze Q-values.