My main line if thought was that I built basically your product for spotting when a cat climbed into my plant pots using ML and a RPi3 - the idea was that when it saw the cat, it would squirt a water pistol at it to scare it away - inference on the RPi 3 was too slow (if I was doing this now I'd use a coral accelerator maybe) and by the time it realised a cat had got into the plant pots, the cat had already taken a shit and left. I worry that your product might suffer from similar end to end latency. Niche use-case? Perhaps. I have Amazon Blink cameras here and the IFTTT integration is delayed by about 30 seconds so by the time you get a notification there is someone at your door it is to late to do anything as they will.have already left/kicked the door in by then etc. Doing all this locally would be super fast

My main concern was not really about privacy - you'll need to cover GRPR if someone from the EU happens to walk into frame of one of your customers' cameras one day in the future anyway (Good luck)