Edge AI · CNN · machine learning
We teach small sensors to see what their datasheet can't.
Most sensors give you a finished verdict — and the verdict fails exactly in the hard cases. We go one level deeper: we read the raw signal and train a compact neural network that runs on the device itself.
Read the raw signal, not the sensor's built-in decision.
Build a labelled dataset on the bench and in the field, including the hard cases.
Train a compact CNN on maps of the signal over time.
Quantize to INT8 — models of a few kilobytes that fit next to the firmware.
Verify on the real device, with measured results in real conditions.
PressureDoors & windows from tiny air-pressure pulses
Openings anywhere in a building, one sensor per floor — now a product: BaroAware.
PyroelectricHuman vs. pet vs. heater
Raw PIR signals classified to cut false alarms.
Time-of-flightPresence from signal coherence
A 4 KB on-device network where distance alone gives up.
60 GHz radarBreathing & presence
Raw radar data for stillness and absence detection.
UWBPositioning in metal-heavy plants
Telling clean line-of-sight from reflections to stabilise indoor tracking.