Workplace Safety Incident Assistant

The problem. After a severe workplace injury, a factory safety officer must report it to OSHA, describe it in OSHA's injury codes (event, source, injury, body part) and decide corrective actions grounded in the OSHA standards. It is slow, the codes are easy to get wrong, and OSHA changed what its codes mean on 1 January 2024. What we built. An assistant that turns a report, typed or read from a PDF form by our Evolus document extractor, into: the four OSHA codes, each with a confidence label measured on unseen 2025 reports (Trusted, Approximate, or Your decision, where the officer must choose before approving); five similar real cases from 17,746 OSHA reports; and 3 to 6 corrective actions ordered by the hierarchy of controls, each citing a 29 CFR 1910 section that may be relevant, with the excerpt one click away. A validator removes unknown citations and legal or medical verdicts. The officer edits, approves and exports; nothing is stored. Interface and drafts in English and Italian. AMD. Every model call runs on AMD Token Factory: Gemma-4-31B-IT gives a second opinion on the event and source codes, and DeepSeek-V4-Flash writes the draft. Evolus (workspace DarkSea28). Our Corrective Action Drafter agent runs on our own Token Factory connection and searches a knowledge library of the 36 OSHA sections; our document extractor reads incoming report forms. The app drives them through the Evolus agent chat and REST APIs, and drafts directly on Token Factory if Evolus is unavailable. Measured, honestly. Tested once on unseen 2025 manufacturing reports: Trusted codes are right 92.8% (event), 96.9% (injury) and 94.5% (body part) of the time; where the app says "Your decision", event codes are wrong 5.1 times as often. Our pre-registered Gemma test failed as registered; we declared a parsing fix and publish both results. Limits. US data, severe injuries only; the drafts are not yet measured with safety officers.