MultiPit

Inspiration

We're huge F1 fans, and we kept wondering: how do teams actually decide when to pit? It's not just tire wear—it's position, competitors, pace, track conditions, all at 200mph. We wanted to build something that makes those split-second strategy calls in real-time, just like the pit wall engineers. Plus, controlling it with hand gestures and sending live updates to an Arduino display like an actual F1 steering wheel? That's just cool.

What it does

Our system is an AI race engineer for Formula 1. It replays historical races and analyzes them lap-by-lap with 4 specialized agents:

Agent	What it watches
Tire Agent	Degradation, predicts when tires "fall off the cliff"
Lap Time Agent	Pace trends, detects when you're losing time
Position Agent	Your position and gaps to nearby cars
Competitor Agent	Rival pit stops, pace, threats

All four feed into a Coordinator Agent (Google Gemini) that decides: Pit now? Stay out? Push or save tires?

Outputs

React dashboard with real-time updates
Arduino LCD showing pit countdowns to the "driver"
Hand gesture controls (swipe to browse strategies, fist pump to confirm)

Multi-Agent Workflow

┌─────────────────────┐
│  FastF1 Race Data   │
└──────────┬──────────┘
           │
           ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│                        Every Lap (Python Agents)                            │
│                                                                             │
│  ┌───────────────┐  ┌────────────────┐  ┌────────────────┐  ┌───────────┐ │
│  │  Tire Agent   │  │ Lap Time Agent │  │ Position Agent │  │ Competitor│ │
│  │  - Deg model  │  │ - Pace trends  │  │ - Gap tracking │  │   Agent   │ │
│  │  - Cliff pred │  │ - Time loss    │  │ - Track pos    │  │ - Rival   │ │
│  └───────┬───────┘  └────────┬───────┘  └────────┬───────┘  └─────┬─────┘ │
└──────────┼──────────────────┼────────────────────┼─────────────────┼───────┘
           │                  │                    │                 │
           └──────────────────┴────────────────────┴─────────────────┘
                                       │
                                       ▼
                          ┌─────────────────────────┐
                          │   Event Detector        │
                          │   Critical moments only │
                          └────────────┬────────────┘
                                       │
                                       ▼
                          ┌─────────────────────────┐
                          │  Coordinator (Gemini)   │
                          │  - Synthesize data      │
                          │  - Pit decision         │
                          └────────────┬────────────┘
                                       │
                   ┌───────────────────┴───────────────────┐
                   │                                       │
                   ▼                                       ▼
        ┌──────────────────┐                   ┌─────────────────────┐
        │  React Dashboard │                   │  Arduino Display    │
        │  - Live updates  │                   │  - Pit countdown    │
        │  - Strategy viz  │                   │  - BOX BOX alerts   │
        └──────────────────┘                   └─────────────────────┘
                   ▲
                   │
        ┌──────────────────┐
        │ Gesture Control  │
        │  (MediaPipe)     │
        └──────────────────┘

Key Innovation: 85% fewer API calls by only triggering Gemini on critical events (78 → 8-12 per race)

How we built it

Tech Stack: Python + FastF1 API + Supabase + Google Gemini + React + Vite + Tailwind + Arduino Uno + MediaPipe

Smart Event Triggering: Four Python agents run every lap (free). Only critical moments trigger expensive Gemini calls—pit stops, tire cliff, position changes.

Challenges we ran into

API costs - Built event-based triggering to stay within free tier
Tire modeling - Ported quadratic degradation formulas from F1 engineering papers
Real-time sync - Buffering between 2s frontend polls and faster backend processing
Gesture tracking - Added cooldowns to prevent false triggers in poor lighting
Arduino serial - Implemented auto-reconnect for USB disconnections

Accomplishments that we're proud of

85% API reduction with smart event triggering
100% pit prediction accuracy on Bahrain 2024 (within 3 laps)
Full-stack in 36 hours: AI backend + React dashboard + Arduino hardware + gestures
Predicted Leclerc's tire cliff in Monaco 2024 (lap 73, actual pit lap 74)

What we learned

Multi-agent systems are more efficient than monolithic AI
Event-driven architecture saves money and compute
F1 strategy is incredibly complex—undercuts, overcuts, traffic, fuel, tire deltas
Hardware integration requires careful handling of serial communication and reconnection logic
Gesture UX needs polish—cooldowns, feedback, forgiving thresholds