Where every number on your label comes fromcomes from

1. The data sources

2. What the AI does — and does not do

AI reads your recipe and works out which ingredient you meant. That is a language problem, and language models are good at it. It never invents nutrition numbers. Those come from the databases above, and all the arithmetic is done in code, where it can be audited and re-run.

When an ingredient cannot be matched with confidence, we do not substitute something similar. It is excluded from the totals and shown to you plainly. If unidentified ingredients exceed 15% of the raw weight, the label is blocked rather than published, and no credit is consumed.

3. The yield and moisture model

Cooking mostly changes water, not solids. So instead of multiplying everything by a blanket yield factor, we conserve the solids and model the water:

• Solids conservation. Protein, fat, carbohydrate, fibre and ash are carried through cooking; water is added or driven off according to the cooking method, and per-100 g values are recomputed on the finished weight.
• Finished batch weight. If you tell us what the batch actually weighed after cooking, that measured weight overrides our model. Your scale beats our estimate every time.
• Frying oil is a bath, not an ingredient. Oil you fry in is not all eaten. We apply a published absorption fraction rather than counting the whole quantity into the product.
• Identical ingredients are merged and evaporated water is never listed in the ingredient declaration.

4. The checks every label must pass

• Mass balance. The declared macronutrients plus water and ash must reconcile with the finished weight. A label that fails this is blocked, not printed.
• Atwater energy. Declared energy must agree with the energy implied by the macronutrients (4/4/9 kcal per g, 2 for fibre, 7 for alcohol) within tolerance.
• Energy ceiling. No food can exceed the energy density of pure fat. Anything that does is a data error by definition.
• Mandatory-nutrient gate. If an FSSAI-mandatory declaration is missing, export is blocked and we name the missing nutrient.
• Regression harness. Every change to the pipeline is re-run against a locked set of previously generated labels. If the pass rate drops, the change is reverted automatically.

5. Two paths, two levels of certainty

6. Honest limitations

• Composition tables report averages. Real ingredients vary by variety, season, soil and supplier.
• Cooking loss depends on your equipment, batch size and time. Our model is published-average behaviour, not your kitchen.
• Micronutrient coverage is thinner than macronutrient coverage in both IFCT and USDA. Where a value is not published, we leave it off rather than fill it in.
• We produce FSSAI-format labels only. US FDA and EU 1169/2011 panels are in development and not available today.
• We are a software tool, not a laboratory, a certifier or a regulator. Nothing here is legal advice, and responsibility for what appears on your pack remains with you as the food business operator.