What's 4 Dinner?

The gap between a useful little utility and something someone pays for every month is wider than it looks. Version one answered a single question well and then sat there. The obvious move was to add AI, and the obvious way to add AI was to ship every user's dining history off to somebody else's server. That would have quietly broken a promise the app had already made in writing on its own store page: your data lives on your device, never sold, never shared. Most apps resolve that tension by updating the privacy policy. We wanted to resolve it by not creating the problem.

We built the intelligence on-device first, using Apple's Foundation Models framework, and treated the cloud as a fallback rather than a foundation. Everything the AI knows comes from a single context document that is assembled on demand from local data and never stored anywhere. That same document became the consent screen: on hardware that cannot run the model locally, the app shows the user the exact text that would leave their phone, verbatim, before any of it does. Consent is per feature and revocable, the fallback server is stateless and logs nothing, and every answer in the app is labeled with where it came from. The pricing followed the architecture. Because on-device inference is free to run, the only tier that costs us anything per request sits behind the tier that pays for it.