At the end of the month payroll has a number for a crew member, and the crew member says it is wrong. The only record of what was said on site is a foreman’s memory and a notebook. CrewVoice lets the foreman log hours by speaking, in English or Spanish, and keeps a record that can be audited.
Nothing reaches payroll unconfirmed. The model can only create a pending entry; the agent reads it back and the foreman says yes before it counts.
Unknown names come back as a question, not as a guess at the closest worker.
Impossible hours are rejected by the tool, so a misheard “20 hours” cannot become a shift.
Every confirmed row keeps what the recognizer heard, next to what was recorded.
Built with
The AssemblyAI Voice Agent API with a FastAPI bridge, deployed on Render. Built for the AssemblyAI Voice Agent hackathon on lablab.ai (certificate of completion).