Your project sounds really interesting. Machine learning models that can be updated and treated more like packages will really give people an edge over people who can't perform that task.

I made a project called mms2concept. What it would do is given a picture taken by a phone (or anything that could send an MMS really) and it would attempt to identify the image based on the picture. I used Clarfai and Twilio to do this.

I think with very resource constrained edge devices either you would have a much more powerful "basestation" style deployments where you have an arduino or similar device collecting edge data and sending it back to that "basestation".