[butterbox]

I am a techno-skeptic most of the time, but I don't think this idea is particularly crazy. Usability-wise, something you can talk to (in your own "low resource" language) is a winner.

The catch is that you have users without either the sophistication to size up how trustworthy this model is or to do out-of-band verification (remember: no internet). I think a RAG model serving answers from Wikipedia could be suitably safe, but we'd have to do some user research to know more.