I don't route all my system's traffic through Gluetun, my threat model doesn't need it, I just route relevant apps, e.g. package management is in the clear but firefox, SearXNG, and nicotine go via gluetun. SSH can look after itself, or I'm in dire trouble. If my threat model did need it, I'd be considering a similar solution to yours, but it'd be heavily cribbed from the known good of gluetun, basically the docker (podman) put to bare metal.
This sounds like thinking I had as carer for my mother with dementia/Alzheimer's. If so, I found introducing new things to be not useful, it will be soon forgotten, (If not, good, and ignore this, but XY-Problem indeed). What I found actually useful was remote video monitoring on my phone (using a few RasPi's, I'm sure there are better solutions, but I was poor, and it worked) allowing me to go out in necessity. The other thing I wish I'd done earlier is reach out to any and all available support earlier rather than later, it's a marathon and every edge you can get will allow you to do better for all concerned, not least yourself.
That said, watches with automatic fall monitoring, with SMS alert to you would be useful. Ideally, something like Alexa (but preferably not as privacy invading, still, needs must as the devil drives) or whatever allowing her to just call out "help" and get a call to you on a speaker / microphone is probably achievable now, perhaps someone can chime in, or you can go searching. Best of luck!
In what way are you not re-inventing the gluetun wheel ? Not trying to put you down, just that I'd need a good reason to consider anything less battle hardened.
Probably unenforceable, like so much of EULAs, but enough to deter the small guys who can't afford the lawyers to defend themselves. Bully tactics. Shame, because an open playing field would benefit everyone but nvidia, also a shame that AMD, who probably could defend themselves, dropped financial support for ZLUDA.
I see this a lot, but do you really think the big players haven't backed up the pre-22 datasets? Also, synthetic (LLM generated) data is routinely used in fine tuning to good effect, it's likely that architectures exist that can happily do primary training on synthetic as well.
