Kitchen Sentinel

Food-safety compliance is normally judged on a single point-in-time inspection, which misses the slow drift toward violations that actually causes most real incidents.

Built for the Google AI House "AI Agent Builder Series 2026" competition (deadline August 5, 2026), after evaluating and discarding a weaker initial concept — a Municipal Health Inspection Copilot — in favor of a kitchen-level compliance agent.

Grounded in real public inspection datasets from NYC and Chicago, used to define what an actual violation pattern looks like rather than inventing synthetic rules.

Four cooperating agents — Vision, Speech, Compliance, and Risk/Report — communicate over MCP and A2A protocols, splitting the problem into perception (seeing and hearing kitchen activity), rule-checking, and risk synthesis instead of one monolithic model.

The centerpiece of the demo is a longitudinal before/after audit narrative: showing compliance drifting over time and being caught, not just a single inspection snapshot.

Coordinating specialized agents decreased response latency and improved compliance accuracy over a single monolithic model approach.

Deliberately moved away from a copilot that assists a human inspector toward an agent that runs continuously — the earlier concept was discarded specifically because it only helped at the moment of inspection, not in the gap between inspections.

Developed full functional multi-agent simulation with comprehensive audit reporting.

Trained and grounded on U.S. public inspection data (NYC/Chicago), so rule definitions would need re-grounding for other regulatory regimes before real deployment.

That splitting perception from judgment across dedicated agents made the compliance logic far easier to explain than a single end-to-end model would have been.