SailSight

ML-powered analytics, visualization, and chatbot system for SailGP racing data. Predicts win probability per team, models dirty air effects, explains predictions with feature contributions, and visualizes races in a real-time F1-style dashboard.

Future work: We intend to integrate SailGP's existing VR visualization system for immersive 3D race replay.

Architecture

/model_training/
/app/
  backend/
  frontend/
/data/Bermuda/

Two separate systems: Training produces model.pkl. The web app loads it for inference only - no training in the frontend.

Quick Start (Local)

1. Train the model

cd model_training
pip install -r requirements.txt
python train.py
# Output: model_training/output/model.pkl

2. Start the backend

cd app/backend
pip install -r requirements.txt
cp .env.example .env   # Add GOOGLE_AI_API_KEY for chatbot
python -m uvicorn main:app --reload --port 8000

3. Start the frontend

cd app/frontend
npm install
npm run dev
# Open http://localhost:5173

ML Model

• Algorithm: XGBoost with probability calibration
• Features: 29 telemetry features across 4 tiers + dirty air
• Output: Win probability (0–1) per team, normalized across fleet

Feature Tiers

1. Positioning - rank, distance to leader/boat ahead, VMG, leg progress
2. Wind - TWS, TWA, TWD, AWS, AWA, wind alignment
3. Foiling - ride heights, average ride height, variance
4. Stability - heel, pitch, yaw/roll/pitch rates
5. Dirty Air - upwind boat distance, relative angle, composite dirty air score

Dirty Air Model

For each boat, boats ahead within 50–200m upwind contribute:

dirty_air_score = Σ(distance_weight × wind_alignment_weight)

This affects speed prediction, win probability, and is visualized as a red overlay on the race map.

Dashboard Features

• 2D Race Map - boats, trails, course marks, wind arrow, dirty air overlay
• Win Probability Chart - horizontal bar chart per team
• Feature Breakdown - explainability panel with positives/negatives
• Chatbot - floating popup (Google AI Studio + local fallback)
• Change Race - switch between Race_1 … Race_8
• Playback - scrub timeline or auto-play race replay

Video Demo

Link

Data

Source: SailGP Data Challenge / Hackathon 2026

Dataset: /data/Bermuda/ - 8 races, 1 Hz telemetry, marks, XML course definitions.

The app reads data_dictionary.md for column interpretation and builds feature mappings dynamically.

Project Structure

model_training/
  train.py              # Main training script
  features.py           # Feature engineering
  dirty_air.py          # Dirty air model
  data_loader.py        # Dataset loading
  output/model.pkl      # Trained model artifact

app/backend/
  main.py               # FastAPI app
  services/
    model_service.py    # Inference + explainability
    data_service.py     # Race state
    chat_service.py     # Google AI chatbot
    dirty_air.py        # Dirty air (inference)
    features.py         # Feature engineering (inference)

app/frontend/
  src/
    App.tsx             # Main dashboard layout
    components/
      RaceMap.tsx       # 2D canvas visualization
      WinProbabilityChart.tsx
      FeatureBreakdown.tsx
      Chatbot.tsx

License

Built for SailGP Data Challenge / Hackathon 2026.

Data © SailGP. Used under the terms of the competition.