ADR-003: Gemma 4 as Edge LLM on Pixel 10 TPU

Status

Accepted

Origin

New for sprint. Replaces Pitwall's hardcoded rules engine (hot path) with an on-device LLM.

Context

Pitwall's hot path uses a hardcoded DECISION_MATRIX — an array of conditions evaluated per frame. This works but has limitations: it can't explain physics, can't adapt language to skill level, and can't reason about context beyond the current frame.

The Pixel 10 has an on-device TPU capable of running Gemma 4 at <50ms inference. This makes it feasible to replace hardcoded rules with an LLM that: - Evaluates telemetry against pedagogical vectors (Ross Bentley curriculum) - Generates contextual coaching, not just threshold alerts - Adapts language to driver skill level (beginner vs pro) - Understands why a technique matters (weight transfer, traction circle), not just when to cue it

Decision

The hot path runs Gemma 4 on the Pixel 10 TPU. Every fused telemetry frame is evaluated by the LLM with matched pedagogical vectors as context.

Inference budget: <50ms per frame. At 50Hz, this means the LLM evaluates every frame if inference is fast enough, or every 2nd-3rd frame if needed.

Pre-filter: A hard-coded confidence gate runs before the LLM. Frames with all critical signals below threshold are not sent to Gemma 4 (saves TPU cycles). This is the only remaining hardcoded rule — everything else is LLM-driven.

Fallback: If the Pixel 10 TPU is overloaded (thermal throttling, background tasks), the system falls back to the Pitwall-style hardcoded decision matrix. The driver gets simpler coaching but never silence.

Prompt includes: - Current frame with confidence metadata - Matched pedagogical vectors (pre-selected by telemetry trigger conditions) - Driver skill level - Last 5 coaching messages (to avoid repetition) - Corner context (approaching/in/exiting which corner)

Consequences

Positive: Coaching is richer, more contextual, and adapts to skill level. The system can explain physics, not just issue commands. Negative: LLM inference is less deterministic than rules. Same frame may produce slightly different coaching. Mitigated by consistent prompting and pedagogical vector grounding. Fallback to rules if TPU is unavailable.