EthnoCAL

Every food tracker I tried misunderstood the meals I grew up eating. Lentejas con arroz became lentil soup; ají de gallina became chicken stew. Before I could log dinner, I had to translate it into something the app understood. That is why I built EthnoCAL to recognize meals the way people actually eat them, not just as a disconnected collection of ingredients.

I also questioned the industry’s obsession with absolute precision. A consistent estimate is infinitely more useful than a theoretically perfect number that requires so much effort people eventually quit. The goal wasn’t another calorie calculator, but a tool you can comfortably use every day without explaining your culture to it first.

The first version of EthnoCAL was photo-first, much like other nutrition apps. It functioned, but I wasn’t convinced it was the right experience. So I built a chat-first interface alongside it, and eventually chose the conversation. Logging should feel like telling someone about dinner rather than filling out a cold, multi-step interface form.

Instead of navigating rigid forms, you simply tell Eth what you ate in English or Spanish. The assistant dynamically parses the text to create a contextual meal card complete with a clear calorie estimate, portion sizes, and a transparent list of any underlying assumptions it made.

One of the biggest decisions was figuring out what AI should actually handle. EthnoCAL uses a catalog for familiar dishes, with Claude interpreting less predictable conversations. The catalog, not the model, determines nutritional estimates. The system must be useful without trusting it to make every product decision.

Testing also exposed severe edge cases. A phrase like me corté abriendo una lata could trigger an incorrect crisis response, while an early test miscalculated a meal by counting the steak twice. I hardcoded sensitive responses using predefined language and wrote regression tests to ensure those critical logical errors won’t return.