A compound AI system that decides go/no-go between motorcycle racetrack sessions. It synthesizes telemetry, ML risk modeling, real-time weather, and historical setup notes into a single trusted recommendation.
At 160 mph, hallucination is a disqualifier, not a bug.
The architecture
Four layers, each picked for what it's actually good at:
- Deterministic Python guardrails. Hard safety limits. The first thing every recommendation passes through.
- Random Forest risk classifier (Scikit-Learn) on telemetry. Real numbers, real probabilities. Not a vibe.
- Real-time weather (Open-Meteo API). Track surface, dew point, wind. The same conditions that change a motorcycle's behavior between sessions.
- RAG over setup notes (ChromaDB). What worked last time at this track, in similar weather, with this tire compound.
Llama-3.3-70b on Groq sits over all of it, but as semantic router and explainer, not as decider. The LLM rarely makes the call. Deterministic rules and ML do that.
Why this matters
I wrote up the architecture and broader thesis as a long-form LinkedIn post: Beyond Chatbots: Why I Built a Compound AI System to Trust My Life at 160mph. The argument: in safety-critical domains (fintech, healthcare, regulated decisioning), single-LLM architectures are the wrong shape. Compound systems that treat the LLM as one component among several let you reason about each component's failure modes independently.
The patterns from AI Crew Chief are now applied to AI-native decision intelligence at SAS. That's the operator-builder loop in action: ship it under the LLC, learn its constraints, port the patterns to enterprise scale.
Roadmap
Active: integrate live biometric data (heart rate, breathing) from Garmin so the system can read rider readiness alongside the bike's.
What this demonstrates
Hands-on LLM and compound AI engineering by a PM. A published thesis in safety-critical AI, not just slides. Direct translation of personal-project design patterns into enterprise decision intelligence work.