Twenty-two menus in, one schema out

Entry 24-1 decided Picky’s schema must survive any client’s taxonomy. This is how you actually build such a schema: from evidence, not imagination.

Julia collected 22 real menus, deliberately weird ones, hotels, restaurants, hospitals, cafeterias, bars ▸ 5:16. Each PDF went into its own Gemini 2.5 Pro chat with one instruction: convert this menu to JSON ▸ 5:52. Being multimodal, it read the allergen icons, not just the text ▸ 6:28. Then all 22 JSONs went back into Gemini together with the framework idea, and out came one master structure robust across every format in the corpus ▸ 7:24. It even suggested features they hadn’t thought of, split-the-bill among them ▸ 8:29.

22 menús reales > 1 esquema imaginado →

Then the beautiful plan hit its own user. The original idea, “restaurants just send us a JSON file”, collapsed under one question: who is going to write a JSON from memory? Picture the mid-size family restaurant. What’s a JSON? ▸ 12:49 (“Jason is a cousin of mine. Greetings, cousin.”) The population problem, how a non-technical owner gets 60 dishes into the system, was promoted on the spot to the most critical part of the product ▸ 14:06. Leading option: they upload a photo of their menu, the AI matches it into the structure, they fill the holes ▸ 14:44.

The rest was filtering. AR dish previews, video menus, Chipotle-grade customization: real ideas, all deferred, “if we start on augmented reality now, paila, we never finish” ▸ 21:20. The MVP reduced to three keys, said in one breath: how the data gets populated, a simple visual layer, and an architecture robust enough to grow into everything else ▸ 21:30…