On May 23, 2026, fresh off the stage at Google I/O, our Google Developer Experts (GDEs) converged on Sonoma Raceway to get inspired and build a real-time, AI-powered race coach. While our pilot proved we could process basic telemetry, Sonoma was about taking the next step: using Antigravity and Gemini to create an AI tool that gives drivers split-second, actionable advice to improve their lap times on the track.

We are closing the AI trust gap by grounding our architecture in physics and real-time verification so people feel confident handing over high-stakes decisions to generative models. For instance, rather than offering theoretical advice, the system pinpointed a new throttle application zone mid-corner in Turn 2, securing a 0.1-second advantage where failure is not an option.

Bridging the domain gap with Google Antigravity

One of the most powerful revelations from Sonoma was how Antigravity served as a domain-bridging engine to build an AI Race Coach that grounded our Trustable AI architecture. Our GDEs, who are expert software engineers, used Antigravity to handle stateful orchestration and telemetry ingestion from the race cars. This allowed builders to focus on high-level system behavior and coaching methods provided by racing experts, demonstrating how AI can empower teams to build real-world applications in unfamiliar domains. The Antigravity product teams were on-site filming this exact transition, capturing how developers move from vibe coding to production-grade deployment at the edge.

From prototype to production-grade Edge AI

At Sonoma, the GDEs operated in a specialized matrix, similar to enterprise adoption tiers to prove Trustable AI scales to meet organizational challenges.

• The Entry-Level Tier (Beginner Pod): This team mirrored the initial adoption phase of enterprise AI, focusing on accessibility and intuitive coaching pedagogy.
• The Optimization Tier (Intermediate Pod): Representing growth-phase integration, they used advanced data logging systems to maximize platform capabilities through precise threshold management.
• The Mission-Critical Tier (Pro-Tier Pod): This team tackled extreme domain gaps in racing, processing massive real-time telemetry and collaborating with pro-level drivers to identify performance gains beyond human perception.

The stack behind the speed

The success at Sonoma Raceway was underpinned by a sophisticated technology stack designed for high-velocity inference. While Antigravity acted as the critical orchestration "glue," the framework leveraged the power of Google Cloud Platform (GCP) and Agent Development Kit (ADK) to facilitate deep, online analysis of telemetry data. This combination allowed the GDEs to bridge the gap between raw data ingestion and actionable strategic insights.

Core product stack:

• Antigravity for real-time code iteration and domain expertise bridging.
• Python for backend telemetry ingestion and data parsing.
• Agent Development Kit (ADK) managing a collection of agents orchestrating key functions.
• Jetpack Compose powering the high-refresh Android cockpit dashboard without dropping frames under heavy telemetry loads.
• Gemini API managing complex post-session driver modeling and cloud reasoning.
• Gemma 4 running locally as an edge-intelligence layer for zero-latency, offline audio coaching alerts.
• Text-to-Speech (TTS) integration for real-time auditory delivery.

Architecture flow:

The architecture flow details a high-velocity data pipeline that translates raw racing telemetry into real-time driver coaching across five key stages. It demonstrates how data moves from edge ingestion and real-time processing to hybrid edge-cloud reasoning, culminating in immediate auditory and visual insights.

1. Ingestion: Python scripts running on Pixel 10 interface with the vehicle to capture real time telemetry directly at the mobile edge.
2. Processing: The Android app built using Jetpack Compose to fluidly track spatial metrics and display corner phases in real-time.
3. Edge Reasoning: Gemma 4 processes the localized stream with low-latency, acting as a fail-safe for real-time alerts when cellular connectivity drops.
4. Cloud Reasoning: Telemetry is synced to the cloud where the Gemini API evaluates performance against our established driver models.
5. Delivery: Immediate insights trigger TTS audio coaching, while complex visualizations update the Compose dashboard.

Our edge architecture required a resilient data pipeline to actually work in the hostile environment of a race car at 100 mph. We are incredibly grateful to community member Brian Luc for solving the hardware gap. He engineered a custom USB interface that wired the Pixel 10 directly into the vehicle telemetry network. This allowed the phone to bypass standard wireless latency and pull a 10 Hz data stream straight from the car’s hundreds of sensors, giving the AI the exact physical inputs needed to execute coaching decisions in real time.

The breakthrough of the Sonoma test was the technical activation of the Pixel 10 TPU. By collaborating with Android engineers to activate the on-device TPU, performance surged to 40 tokens per second. This jump provided the real-time reliability required to deliver coaching exactly when the driver needed it.

This architecture translates directly to mission-critical enterprise domains. Startup founders like Vijay Vivekanand (COI Energy) and Jorge Mendieta (Bloom Energy) joined the cohort to explore how agentic orchestration can secure energy pipelines and manage agriculture respectively. By proving the framework at 100 mph, we are paving the way for trustable AI in industries where failure is not an option.

The global sprint: Destination interlagos

The Sonoma evolution is just the beginning. To maintain our momentum, the initiative heads next to Interlagos, Brazil. There, we will further harden the architecture in a new climate and complex track configuration, continuing our mission to bridge the AI Trust Gap across the world.

Members of the July cohort of the "AI Field Test"

Ready to build?

Get hands-on with the same tech we used on the track. If you are ready to move beyond vibe coding and start building on the pro-tier of Vertex AI, get started with our ADK Crash Course. Then, take it to the next level by building your own AI Race Coach with the Trustable AI Codelab.

Deep Dives from our GDEs

• ApexAI Implementation by Henry Ruiz: Repo for Team 3's pro-tier edge ML.
• Trust the Curb, Trust the Commit by Taha Boushine: Building production-grade AI agents.
• Driven by Data, Trusted at Speed by Simon Margolis: Lessons in technical reliability.
• Racecraft: The Origin Story by Rabimba Karanjai: Navigating Project Koru.
• Beyond the Track by David McLaughlin: The vision for Trustable AI.