Build log: what happens when you snap a meal

This is a build log, part of a series on how Snapeto actually works. We publish these because food apps have earned distrust by hiding their mechanics behind confident numbers, and because we've already written about what cameras can and cannot know. Sunlight is the policy.

Step one: recognition, in your language

The photo goes to a vision model that answers three questions: what dish is this, what are its visible components, and what rough scale does the plate suggest. Recognition is the strong link in the chain, and a one line hint makes it stronger. In Snapeto that hint can be typed or spoken in more than a dozen languages, including Indic languages the big Western apps never bothered with. "Ghar ka rajma chawal" resolves the way it should, because the model hears the dish the way you'd actually say it.

You always see what the model concluded, and you can correct it in a tap. That correction loop matters more than any benchmark: a wrong guess silently accepted would poison everything downstream, and every correction teaches us where recognition needs work on real dinners in real kitchens.

Step two: estimates that admit they are estimates

From the recognised dish and portion class, Snapeto estimates calories and macros. We show them, and we label them for what they are: Ai estimates, not medical advice. The research is unambiguous about what a photo can support, portion estimation errors around 200 calories per meal are normal even for automated systems, so we render ranges of confidence honestly instead of dressing a guess in laboratory clothes. The estimate's job isn't to be a verdict. It's to be good enough to steer the week, and for that job it is genuinely good enough.

Step three: the healthy twin

This is the step that makes the snap worth taking. Alongside the estimate, the model generates a healthier twin of the same dish: your korma with yogurt carrying the sauce, your fried rice with more egg and less oil, tuned to your goals, your allergies, and what your pantry actually holds. The information a tracker would have buried in a graph arrives instead as something you can cook on Thursday. Data that looks backwards, converted to a decision that looks forwards, in one generation step.

Step four: the week check

Finally the meal updates your week. Your guardrails are behavioural, cooked dinners, protein presence, rhythm, so the check is behavioural too: on plan today, or a gentle steer pointed at the next meal rather than a debt hung on the last one. No day is graded. No streak exists to break. The week bends, absorbs, continues.

What we watch, on the engineering side

• Correction rate. How often users fix the dish guess, split by cuisine. This is our real quality metric, and the reason regional dishes get first class attention.
• Hint usage. Meals snapped with a text or voice hint resolve dramatically cleaner. Lesson: make the hint effortless, never mandatory.
• Time to done. The whole loop has to beat typing a search into a database, or it loses to the old chore on the old chore's terms.