Keeping the AI on the device

Jul 2026 · Shaaf Kashan

Mihwar is a personal-wellness app I built where the AI companion runs on the phone itself: a small language model doing its work locally, with no servers on my side at all. A person's reflections are about the most private thing they write down, and they never leave the device.

On-device AI is more constrained than calling a big model in the cloud. You live inside the phone's memory and compute; there's no data-centre GPU to lean on. For anything handling sensitive data, though, that constraint is the whole point. No server log to leak, no third party to trust, no per-message bill that climbs with every new user.

Plenty of AI products send everything off to an API by default and treat privacy as a page in the settings. I'd rather make it a property of the architecture. If the data physically never leaves the device, you don't have to promise you'll look after it.

It's the same instinct as the last piece: decide where a capability should live before you build it, not after.