PitSense Pro

AI-Powered Racing Strategy for the Toyota GR Cup Championship

Comment

Inspiration

As a motorsports enthusiast and being passionate about ml and ai, I noticed a critical gap in racing strategy. While teams collect terabytes of telemetry data, most pit stop decisions still rely on crew chief intuition rather than real-time AI analysis. I was inspired by the Toyota GR Cup championship where split-second decisions determine race outcomes. What if we could combine modern machine learning with racing expertise to create an AI co-pilot for pit strategy?

What it does

PitSense Pro is an AI-powered racing analytics platform that transforms raw telemetry data into actionable pit strategy recommendations. Our system:

  • Predicts tire degradation across different track conditions using ensemble learning
  • Integrates live weather data to adjust strategy in real-time
  • Analyzes 3,107 laps across 6 GR Cup circuits for track-specific insights
  • Provides what-if simulations for different pit stop scenarios
  • Offers sector-by-sector performance predictions using custom ML models

How I built it

Tech Stack: Python, Streamlit, Scikit-learn, Plotly, Hugging Face Datasets(for data storage and deployment)

Data Pipeline:

  1. Ingested 2GB+ of racing data including lap times, telemetry, and race analysis
  2. Built universal track models that adapt to any circuit's characteristics
  3. Implemented ensemble forecasting with Random Forest and time-series analysis
  4. Created real-time weather integration using OpenWeather API
  5. Developed interactive visualizations for strategy comparison

Key Innovation: My UniversalTrackModel class can generate accurate predictions for any track, even with limited historical data, by understanding fundamental racing dynamics.

Challenges I ran into

  • Data heterogeneity: Different tracks had completely different data formats and telemetry parameters
  • Time parsing nightmares: Lap times came in 5+ different formats (MM:SS.sss, total seconds, mixed formats)
  • Model generalization: Creating AI that works across diverse tracks like technical Sonoma vs high-speed COTA
  • Deployment constraints: Managing 2GB+ datasets within platform storage limits
  • Domain knowledge gap: Learning racing-specific concepts like tire compounds, degradation curves, and pit window strategies

Accomplishments that I am proud of

  • Built working AI models that accurately predict pit windows within 2-3 lap accuracy
  • Processed 3,107 real racing laps into actionable insights
  • Created universal architecture that works across 6 completely different tracks
  • Achieved real-time performance with complex ML models in production
  • Designed intuitive UI that makes advanced analytics accessible to non-technical users

What I learned

  • Racing is fundamentally about energy management - tires, fuel, and mechanical wear
  • Track characteristics dramatically affect strategy - what works at Barber fails at Sebring
  • The "undercut" is mathematical - we can quantify the lap time advantage
  • Data quality beats algorithm complexity - clean telemetry is worth 1000 sophisticated models
  • Racing teams need explainable AI - they trust recommendations they understand

What's next for PitSense Pro

  • Real-time telemetry integration with live data feeds during races
  • Multi-driver strategy optimization for team racing scenarios
  • Advanced weather modeling with radar integration and precipitation forecasts
  • Mobile app for pit wall crew with push notifications
  • Expansion to other racing series - NASCAR, F1, endurance racing
  • Partner with racing teams for real-world validation and improvement

Built With

Share this project:

Updates