Solving for Latency and Connectivity in High-Stakes Environments
The core challenge of building an AI race coach is the requirement for near-zero latency and reliable operation in environments without internet connectivity. While cloud-based models like Gemini are powerful, they are unsuitable for real-time track coaching due to the dependency on cloud access. To solve this, the team utilized Google's open-source Gemma 4 model, fine-tuned with specific guardrails and deployed locally on a Pixel 10 device. By leveraging the device's onboard TPU, the system achieves the necessary speed to provide coaching feedback at 100 mph without needing a network connection.
Agent-First Development and Rapid Iteration
The project utilized Google Antigravity, an agent-first development platform, to bridge the gap between complex hardware telemetry and application logic. The system extracts raw vehicle data—including GPS, accelerometer, pedal position, and steering angle—via a CAN-to-USB converter.
Key takeaways from the development process include:
- Building on Production: The team treated the racetrack as their development environment, iterating on the AI's behavior in 20-minute windows. When the initial model proved too "chatty" or misaligned with the driver's location, they were able to patch the logic and redeploy immediately.
- Brain Transfer Learning: The system uses a feedback loop where data from the driver and coach are exported as files, which the local LLM then uses for inference to refine future coaching suggestions.
- Abstraction of Complexity: The platform allowed developers to focus on the coaching logic rather than the low-level engineering required to sniff data from the vehicle's CAN bus, significantly accelerating the development cycle.
Trust and Confidence in AI Systems
A critical insight from the project was the evolution of the AI's "confidence" as a coach. Initially, the model struggled with timing and verbosity. Over several iterations, the system learned to provide concise, actionable prompts (e.g., "throttle here") only when necessary. This project serves as a proof-of-concept for "trustable AI," suggesting that if an AI can be trusted to guide a driver in a high-G, high-danger environment, the same architectural patterns can be applied to medical or environmental monitoring where split-second accuracy is equally vital.