Our own product
Say the word, and that
becomes the filter.
psvision is an AI camera that runs on a Bluetooth LE microcontroller. The classifier runs on the sensor itself, so what leaves the device is a verdict: a class, a confidence, and one still of about 70 KB as evidence. No MIPI. No application processor. No video stream.
Built on the Arducam B0642, Sony IMX500
One unbroken take. With the filter set to laptop, a laptop walks into frame and the app reports the laptop, not the person carrying it. Say person, and it finds him.
How it works
Speak
Always on keyword spotting runs on the nRF54 Axon NPU. The word you say becomes the only class the firmware forwards.
See
YOLO11n at 608 by 608 runs on the Sony IMX500 own die, on an Arducam B0642. Inference on the sensor, over plain SPI.
Know
Detections arrive on an Android or iOS phone over Bluetooth LE: the label, the confidence, and the frame that caused it. Kilobytes per event instead of megabytes per second.
The honest boundary
Spec
Field tested
Two conferences, one week, one rig.
Teardown in Portland, three days on a Nordic table in front of makers and developers. Austin Circuit Live five days later, in front of OEM design engineers. The same unit held up in both rooms.
Ultralytics YOLO Vision 2026, Shenzhen, September 13
The unit ran all day at Sony’s booth, on Arducam’s table, without a restart. Visitors tried it themselves: say a word, and the camera reports only that.
Build log
The full write up is live, including what broke: on sensor YOLO11n, keyword spotting on the Axon NPU, and the buffer arithmetic that holds the two together. Firmware, phone apps, and enclosure were built end to end by our team, on Arducam and Nordic hardware.
Want one of these on your product?
On sensor inference, a BLE class host, and an app that makes it a product. That is the PoC Sprint, from the team that built psvision.