An AI food tracker: a photo, a voice note or a barcode, and the data is stored properly

Four ways to log the same dinner

The result is the same record; what changes is what it costs you:

Kodi

(Photo of the plate: grilled salmon with rice and broccoli.)

“Grilled salmon (140 g), cooked rice (180 g) and broccoli (90 g): about 610 kcal, 41 g of protein. Portion estimated from the photo: if the rice was less, tap it and I will adjust.”

🎤 “Had salmon with rice and broccoli for dinner, the salmon was a normal fillet.”

“Logged as dinner: salmon (150 g), cooked rice (150 g), broccoli (100 g). About 590 kcal and 43 g of protein.”

Salmon rice broccoli, about 700 kcal according to the menu

“Logged with your 700 kcal, no recalculation. I am estimating the macros from the dish and flagging them as pending refinement.”

(Scans the barcode on the salmon pack.)

“Smoked salmon, 100 g pack (Open Food Facts). The whole pack or part of it?”

Tracker vs counter: why the difference matters

A counter optimises the day’s number; a tracker optimises the record. In practice that means four things Biokode does that most “calories from a photo” apps do not: it stores where every nutritional value comes from, it marks which part of the entry is an estimate and which is a measurement, it lets you correct without deleting, and it keeps the real time of the meal so it can later be crossed with sleep or training.

With that, today’s entry is still worth something months on: for the pattern engine to compare nights, for the adaptive plan to compute your real expenditure and for micronutrients to accumulate by week against a serious reference.

The four ways in

• Photo: the AI identifies the foods on the plate and estimates each portion. Every estimated portion is flagged as such and corrected in one tap. Median latency in production: 5.9 seconds per photo.
• Voice: you describe the meal the way you would tell someone (“toast with tomato and olive oil and a macchiato”) and the app turns it into foods, quantities and meal occasion. Five to six seconds for one or two dishes.
• Text: the same sentence typed; handy when you cannot talk. If you say “about 700 kcal according to the menu”, those calories rule and are not recalculated.
• Barcode: the product comes from Open Food Facts with its label; you decide the quantity. If you edit the portion, every nutrient rescales in proportion.

Where the numbers come from

Every food in the catalogue has its source: USDA for most composition data, CIQUAL (the French database) to fill what USDA does not cover, and Open Food Facts for packaged products by barcode. Calories are not read from a table: they are computed from the macros (4 kcal per gram of protein and carbohydrate, 9 for fat, 7 for alcohol), so they always add up to what you see.

There are 27 micronutrients, compared with EFSA dietary reference values for your sex and age. Iron, magnesium, vitamin D, folate, zinc, iodine: every week you see your coverage against a European reference, not an invented number.

What the app does when it is not sure

A tracker’s honesty shows in the ambiguous cases, and there are a few. These are the rules Biokode applies:

• Estimated portion: if the quantity comes from a photo or a description without a weight, it is flagged as estimated and Kodi offers to adjust it, without nagging.
• Raw or cooked weight: rice, pasta, legumes and oats are read as cooked weight by default, which is how they look on the plate. If you say “80 g of rice, raw”, that is respected.
• Your calories rule: if you provide the kcal from a menu or a pack, they are not recalculated; macros are estimated and flagged as pending refinement.
• Drinks: a coffee or a tea is logged by the liquid, not the concentrate; water goes to hydration, not to food.
• Correct without duplicating: “it was two servings, not one” or “drop the bread” edits the last entry instead of creating another.
• Meal occasion is inferred: breakfast, lunch or dinner from your usual times; if you say “at 16:00”, that is the time stored.
• Implausible amounts: “2 kg of pasta” is confirmed before it is logged.

How wrong it gets, measured in public

Photo portion estimation is measured on Nutrition5k, a public dataset from Google Research with real dishes photographed from above and the scale weight of every ingredient. With 150 dishes, the same model and prompt as in production, over two runs (August 2026): a median error of 23 to 25% in plate weight, 63 to 65% of dishes with an error under 30%, and a mean bias of ±1%, meaning it neither inflates nor trims systematically.

Two caveats go with the figure. The photos in that benchmark are top-down, with no cutlery and no size reference: harder than a phone photo. And a 25% error in weight is not a 25% error in your day: it averages out across dishes and days, and the portions you correct stop being estimates. That is why the number is published in the methodology with its limits, and not in a headline.

What happens after you log

• The day closes with calories, macros and micronutrients against your plan and your reference.
• The week accumulates micronutrient coverage and the adaptive plan compares your intake with your weight trend.
• Over time, the pattern engine uses your meal times to know whether you ate late and compares that with your sleep.
• Kodi tells you what matters in the morning and at close; in between, it answers when you write. In data-only mode, you read it yourself on the dashboard.

Frequently asked questions

Keep going