The problem we set out to solve
Every gas furnace has three main moving components: the inducer motor, the blower motor, and the ignition system. Each one fails in a predictable pattern. The question is whether you have sensors in place to observe that pattern before it becomes a no-heat call.
Most HVAC companies don’t. They find out a motor is failing when the homeowner calls, usually on the coldest night of the year.
What we built
FurnaceWatch is a small IoT sensor that clips onto a furnace chassis and runs continuously. It samples vibration at up to 2,048 Hz using a precision ADXL355 tri-axis accelerometer, computes FFT features locally, and runs a trained TFLite neural network on an ESP32 microcontroller.
No data leaves the device for inference. The model runs on-chip in under 40ms.
When the model detects an anomaly — bearing wear, blower imbalance, flame sensor degradation — an alert fires through MQTT to the FurnaceWatch platform, which routes it to the right technician within 2 seconds.
The accuracy numbers
After training on 103 labeled furnace cycles across multiple HVAC models:
- 94% detection accuracy on the held-out test set
- <3% false positive rate at standard operating thresholds
- 6 out of 7 failure modes detectable before visible symptoms
What’s next
We’re opening our beta program to a limited number of HVAC service companies. If you manage a fleet of 10 or more residential units, request a demo and we’ll ship evaluation hardware.
The software is already running. The question is whether your team is ready to get ahead of failures instead of responding to them.