PawVeda

the problem

A model alone should not make this call.

A worried owner typing symptoms at midnight does not need a fluent paragraph. They need to know whether this is a vet visit, and a decision like that cannot rest on generated text alone.

what I built

Rules first, model second.

Built with a co-founder. A local rules engine scores the symptoms on the device first, then the model adds context. The local verdict is computed before the model runs and travels attached to the response, so the app can always say what it decided even if the model says nothing at all.

how it stays honest

Anything not real says so on the screen.

Every surface that is mocked, scripted or local only carries a label saying what it is: one chip with eight states, used across thirty two files. Every generated image is logged with the prompt that made it, the model, the date and where it is used.

That discipline includes this page. The capture here is the shipped Android build running on a device, and you can see the demo data chip in the screen itself.