muse-fieldnotes

Fieldnotes turns an HVAC technician's photos, voice notes and texts from a job into the close-out note the office bills from, and the note checks itself. Every line cites the evidence it rests on, a second model checks every line, and when something is missing or doesn't add up, the agent asks the technician on their phone instead of guessing. The problem: every service shop bills from notes like "topped it off with 410A, about two pounds." When something is missing, the office calls the tech, who is already on the next roof. AI note-writers make it worse by filling the gap with a confident guess. In our live test, the tech said 410A, but the rating plate said R-454B. Every line of the note was true to what it cited, so a line-by-line checker would pass it. Fieldnotes caught the conflict. It asked the tech, who answered by voice from their phone with no app and no login, and the note the office signs off now records what actually went into the system. It's live at muse-fieldnotes.vercel.app, with verifier failover across three hosts, tests in CI, and measured accuracy on its own tasks.