Every factory floor already has cameras. The problem is that nobody is watching them. A safety officer can't be on every camera, on every shift, on every floor — so the footage gets recorded and reviewed after someone is already hurt.
The obvious fix — "detect no hardhat → fire an alarm" — is worse than it sounds. It's hardcoded, it can't read your site's rules, and it drowns the ops team in false alarms until they mute it. What's actually missing is something that reasons about risk the way a human EHS officer does — and can prove why.
So I built SafetyCommander: an autonomous EHS agent that watches the floor, reasons about hazards from your written policy, grounds every alert in OSHA law, routes it to the right worker, and hands off a report. It started at the Zapdos Labs × Antler hackathon (AI Agents for the American Industrial Revolution) and I've kept building on it since.
This post is about the one design decision that made it work, and two engineering war-stories that were more interesting than I expected.
The one decision that matters: risk lives in exactly one module
