What it does
KeiVideo is designed to intercept the raw sensor stream and run the inference work — detection, classification, scene understanding — before frames ever leave the camera module. The host application would see a clean structured output instead of paying the cost of decoding and processing full-resolution frames.
Why on-sensor
Bandwidth, latency, and power all favour pushing inference as close to the sensor as possible. The result is a camera that ships features, not pixels.
Roadmap target
Dedicated silicon or an FPGA fabric implementation, in that order of preference — the lab already runs hardware accelerators in an FPGA fabric elsewhere (see the MAX-OS silicon page), which is the natural proving ground before any fixed silicon commitment.
cross-references