GR-GridMind

"In racing, data is noise. Drivers don't need charts; they need answers."

Comment

💡 Inspiration In modern racing, data is abundant, but intelligence is scarce. Drivers are overwhelmed with raw numbers on their dash. We realized that a driver doesn't need to know their "Lateral G-Force is 1.4"; they need to know "The opponent ahead is sliding. ATTACK NOW."

We wanted to build the "Iron Man HUD" for the Toyota GR Cup—a system that moves beyond passive analytics to Active Adversarial Strategy. We asked: "What if the car could think?"

🏎️ What it does GR-GridMind is an Agentic AI Race Engineer that functions as a "Digital Co-Pilot."

The Time Machine: It ingests historical raw telemetry and streams it as a live, millisecond-accurate broadcast.

The Golden Sync: It uses a custom GPS algorithm to align the Hero car and Rival car by track distance, not time, allowing for perfect real-time comparison.

Adversarial Prediction: It calculates an "Overtake Probability Score" (0-100%) in real-time. When the probability spikes, the UI triggers "ATTACK MODE."

RAG Strategy Brain: Utilizing pgvector, the AI recalls historical track advice (e.g., "Blind entry at Turn 1") and displays context-aware strategy on a "Dynamic Island" UI.

⚙️ How we built it We architected the system into three specialized squads:

Squad A (Data & Physics): We used Python & Pandas to build "Self-Healing Data Agents." The raw dataset had a critical bug (ECU Lap Counts reading #32768). Our agent implemented a GPS Geofence Gate to mathematically repair the lap counters and derive physics metrics like Tire Stress and Cornering Efficiency.

Squad B (The Engine): We built a high-performance streaming backend using FastAPI and Redis. We utilized PostgreSQL with pgvector to store both relational race data and vector embeddings of track strategies, enabling Retrieval Augmented Generation (RAG).

Squad C (The Holographic UX): The frontend is built on Next.js 14 and Three.js. We created a 3D Digital Twin of the car that rotates based on live G-Force data and a "Glassmorphism" UI optimized for dark-mode readability.

🚧 Challenges we ran into The "Dirty Data" Nightmare: The provided dataset contained erroneous lap counts (Lap #32768) and unsynchronized timestamps. We had to write a custom "Golden Sync" algorithm to interpolate two cars onto a unified distance grid.

The WebSocket Barrier: Connecting the high-frequency Redis stream (Backend) to the React store (Frontend) initially caused massive CORS and latency issues. We solved this by implementing a Zustand store that updates via direct mutation, bypassing React's render cycle for 60fps performance.

🏆 Accomplishments that we're proud of Successfully implementing "The Hunter Logic"—a mathematical model that predicts overtakes before they happen visually.

Building a "Time Machine" architecture that makes historical CSV data feel like a live, heart-pounding broadcast.

The "Attack Mode" visualization—seeing the screen pulse Red when the math says "Go" is a visceral experience.

🚀 What's next for GR-GridMind Voice Synthesis: Converting the AI's text strategy into a voice that speaks directly into the driver's earpiece.

Multi-Car Tracking: Scaling the "Golden Sync" to track the entire grid, not just one rival.

Built With

Share this project:

Updates