AirSentinel
Fault detection for the air handling units of large buildings. It reads every unit in a fleet, names the fault, points at the part that has failed on a live drawing of the unit, and gives the technician the repair steps.
- A LightGBM model names 12 faults — stuck dampers, passing valves, fouled coils, clogged filters, slipping belts, leaking ducts and drifted sensors
- Scored only on units the model never saw: 97.5% of diagnoses correct and 0.41% false alarms
- A live twin of each unit: air moves at the fan speed, is coloured by its temperature, and the failed part pulses
- Every fault comes with where it is, what it looks like, likely causes, repair steps, tools and time
- Command centre with the fleet drawn as a city, a live count of the money open faults are wasting, and a seven-day time machine
- Ask Sentinel answers questions in plain words from the live fleet data, and can read the briefing aloud
- Fault Lab: break a unit on purpose and watch the model find it
- Printable work order for any open fault
Built with
- Machine learning
- LightGBM, scikit-learn, pandas
- Language AI
- Self-hosted LLM, Web Speech API
- Backend
- Python, Flask
- Frontend
- React, Vite, Framer Motion