How to track mixed meals accurately with AI

Mixed meal: any dish where the ingredients are combined and no label exists for the whole — a curry, a lasagna, a burrito bowl, Tuesday's leftover stir-fry.

Mixed meals are the single biggest reason people quit tracking. A packaged snack takes five seconds to log; a homemade dal with rice and ghee can take five minutes of database spelunking and still end up as a shrug. Here's the method that actually works, and how AI changes its cost.

Why searching "chicken curry" fails

In a crowdsourced database, "chicken curry" returns hundreds of user-typed entries spanning roughly 90 to 250 kcal per 100 g — because "chicken curry" isn't one food. A coconut-cream korma and a tomato-based jalfrezi differ by a factor of two at the same weight. Any single number for the dish-as-a-word is wrong for your version of it.

The truthful unit of a mixed meal is the ingredient. Your curry is: chicken thigh, coconut milk, oil, onion, tomato, rice. Each of those IS one food with one USDA record:

Decomposed, the meal is not a guess — it's a sum of five checkable lines.

What the AI is good at (and what it isn't)

AI-assisted logging works because the two halves of the problem have very different difficulty profiles:

• Recognizing ingredients — what models are genuinely good at. From "leftover dal with rice and a spoon of ghee" or a photo, current models identify components with high reliability.
• Estimating portions — where honest uncertainty lives. Grams from a photo are inherently approximate: depth, bowl size, and density are partly hidden. Expect ±20–30% per ingredient from a photo, better from a text description that states amounts.
• Reciting nutrition numbers — what models are *bad* at. A language model asked for calories directly will produce a plausible-sounding figure that can't be traced anywhere. This is the step that should never be left to the model.

Calorie.one draws the line exactly there: AI proposes the ingredient list and gram amounts; the numbers for each line come from USDA FoodData Central. Every line carries a source label — usda for a direct database match, mixed for a partial match, ai for the rare line that had no acceptable match and is flagged as a model estimate.

The verification step is the accuracy

Because the output is lines, not a total, you can fix precisely the part the AI got wrong. The model read your bowl as 250 g of rice but you know your rice fist is 180 g? Edit that line; the USDA per-100 g basis reprices it instantly. This is the practical difference between "AI calorie counting" as a party trick and as a method: per-line editability turns an estimate into a collaboration. Three corrections you'll make most often:

• Portion sizes of the calorie-dense minority (oil, cheese, nuts, dressing) — small grams, big kcal.
• Cooked vs raw assumptions — cooked rice is ~130 kcal/100 g, dry is ~365; the AI states which it assumed, so check it.
• Hidden fat in cooking — add a line for oil if the dish was fried and the AI didn't see it.

Repeat meals: log once, reuse forever

Most people rotate 15–25 dishes. Once a mixed meal has been decomposed and corrected once, logging it again is instant — and the accuracy compounds, because your correction, not the first guess, is what gets reused. The hard problem shrinks to genuinely new dishes, which are maybe two a week.

FAQ

How accurate is AI photo calorie counting really?

For ingredient identification, very good; for portions, expect roughly ±20–30% per ingredient from a photo alone. Decomposition helps: independent errors on five lines partially cancel, so the meal total is usually tighter than the worst line. A text description with stated amounts is more accurate than a photo.

Why not just let the AI state the calories directly?

Because a language model generates plausible numbers, not looked-up numbers, and there's no way to audit where a generated figure came from. Grounding each ingredient in a USDA record keeps every number traceable — and keeps the model's imagination out of your data.

What about recipes I cook in batches?

Weigh or estimate the whole batch's ingredients once, then log the fraction you eat (a 6-serving lasagna → each serving is one-sixth of each line). The per-ingredient method scales down cleanly because the USDA basis is per 100 g.

Do I need to weigh food for this to work?

No — reference-based estimates (a fist of rice ≈ 150–180 g cooked, a palm of meat ≈ 100–120 g) are enough for a useful weekly average. A scale is an upgrade, not a requirement, and it matters most for calorie-dense foods like oil, nuts and cheese.