Methodology

The AI never invents a number. Here's the whole method.

Most photo calorie apps ask an AI model to output calories directly. Models are good at recognizing food and bad at arithmetic recall — which is why the same plate can return three different numbers. Proba splits the job:

1. The model identifies, it doesn't count. The vision model's only output is a list of foods and estimated gram weights — {food, grams}, nothing else.
2. The database does the counting. Each food resolves against USDA FoodData Central — public-domain, lab-analyzed data. Calories and macros come from the matched USDA entry, scaled by grams. Packaged foods use the barcode label directly.
3. A reconciliation gate rejects nonsense. Every result is checked against the Atwater relation (4·protein + 4·carbs + 9·fat). If the calories deviate more than 20% from what the macros imply, the macros win. Physically impossible database rows (over 900 kcal per 100 g, macro sums over 105 g/100 g) are rejected.
4. You see the receipt. Pull down any result: every line shows its grams and its USDA source. If a food couldn't be grounded, the receipt says so — it never fabricates a citation.
5. Determinism. Same photo → same foods → same USDA rows → same answer. Run our five-scan test on any app, including this one.

What the glucose line is (and isn't)

Each receipt carries an estimated glucose-impact band computed from meal composition (glycemic load from carbs, fiber, fat, protein). It is labeled in-app exactly as what it is: an estimate from meal composition — not a CGM measurement. If you wear an over-the-counter CGM, Proba can read your Apple Health glucose data to pair meals with your real curves.

Honest limitations

Accuracy claims worth trusting are testable. Ours are: the citations are on every receipt, and the repeatability test takes five scans of one meal.