Maintenance Copilot: Catch Failures Early

Factories already collect sensor data, maintenance logs and manuals, but fixed-threshold alarms fire late, when a machine is already close to failure. Technicians then lose hours searching manuals and old work orders while unplanned downtime costs thousands of dollars per hour. Maintenance Copilot is a working web app for maintenance engineers and shift supervisors. It compares every machine with its own 7-day baseline and flags drift early. On our demo plant it caught an M3-Pump bearing fault 39 hours before a fixed 4.5 mm/s vibration alarm, worth about $195,000 of avoidable downtime at $5,000/hour. One click sends the sensor evidence plus the most relevant manual sections and past work orders to Gemma-4-31B-IT on AMD Token Factory. The model returns a structured diagnosis: likely cause, urgency, confidence, numbered actions, a safety note and the exact sources used. The app validates the JSON, retries once if it is malformed, and shows a "human review recommended" banner when confidence is low or urgency is high. A technician chat answers questions like "How do I replace the drive-end bearing safely?" using only the plant's own documents, with sources shown. All inference runs on AMD (about 6-9 s and 1,000 tokens per diagnosis), with live latency and token tracking and graceful handling of rate limits. Built with Python, Streamlit, Altair and a custom TF-IDF retriever; runs on Streamlit Cloud or Docker, and accepts your own sensor CSV.