On-Device Meter Reading AI
In active development
Computer-vision system that reads an electric meter from a photo — both 7-segment LCD and analog clock-face dial meters — running entirely on the phone with no network call. Two-stage YOLO detection locates the meter face and then the register panel; a CRNN + CTC reader decodes the digits. Trained in PyTorch on Apple Silicon, exported to ONNX, and shipped as a Kotlin/CameraX Android library.
Problem solved: Field readings were keyed in by hand with no way to verify them against the photo, and readers work in areas with no signal. The system reads the meter offline on-device and cross-checks the reader’s claimed value, flagging mismatches for review instead of trusting the keystroke. Ongoing R&D — each generation of models is measured against real field photos with a frozen champion as the model to beat, and retrained as new capture areas come in.