Science and evidence

MyMeals isn’t a chatbot guessing at your plate. A vision model describes the photo. A deterministic health engine, written in Rust, applies strict epidemiological rules to score it. MyMeals grades every health measure with the WCRF evidence framework. No black boxes.

Updated September 25, 2026

From photo to meal

1. Describe. A frontier vision model from Anthropic, OpenAI or Google reads the photo, or several photos of the same meal together, and describes each item in general terms: “toast, two slices, with cheese”, “a glass of orange juice, about 250 ml”. It also reads the cooking method and the portion cues in the frame: the plate, the glass, a hand, or a bowl you registered as Personal Tableware. MyMeals keeps receipts and supermarket shelves out of your meals.

2. Match. MyMeals maps each description to a specific food in food-composition data and converts the portion to grams using that food’s own units: a slice, a glass, a tablespoon. The nutrients come from the data, not from the AI model. A barcode or a photographed nutrition label identifies a packaged product exactly and always takes precedence. When no food matches well, MyMeals keeps the item in the meal and flags it to calibrate.

3. Personalize. Your health graph fills in what the photo leaves open: the brand of oat milk you buy, the products on your recent grocery receipts, the amount you usually pour, the bread you eat every week. Clear visual evidence, such as a visible package, always outranks your habits.

4. Score. The health engine runs on your iPhone. It applies the definitions below to the confirmed meal and returns the same result for the same meal every time. The AI never decides a score.

Every meal lands in your inbox as a draft. A two-second review makes it exact.

Scientific honesty, not false precision

A 2D photo cannot reveal hidden butter or exact depth. The oil in the pan and the butter under the potatoes drive most of the error in any photo-based calculation.

Instead of faking precision, MyMeals calculates a strict interval for every uncertain item, such as 170 to 330 kcal for bread of unknown thickness, and carries the combined interval into the meal total. MyMeals flags those items, so calibrating one or two locks in a precise meal.

The interval narrows when:

• you confirm or correct an item, or pick the right alternative;
• a barcode or label identifies the product;
• you have tracked the food and portion before;
• the photo includes your Personal Tableware;
• you weigh the food and enter the weight.

Where the numbers come from

Some attributes have no complete database, such as the glycemic index of every food. MyMeals derives these once per food with an AI model, stores each with a range and a confidence grade, and labels it as AI-derived wherever it appears.

The metrics that drive healthspan

Calories are a unit of heat. They don’t tell you how food behaves in your body. MyMeals tracks the metabolic context of every meal.

Microbiome diversity

MyMeals counts the distinct plant foods in your meals (vegetables, fruit, whole grains, legumes, nuts and seeds) and keeps a weekly tally toward 30. In the American Gut Project, people who ate more than 30 plant types a week had more diverse gut microbiomes than those who ate 10 or fewer [1]. Microbiome diversity is an association, not a guarantee of health, and the evidence grade in the app says so.

Metabolic load (NOVA)

Every food carries a NOVA class, from unprocessed to ultra-processed. MyMeals takes it from product data where it exists and otherwise infers it from ingredients and additives with a fixed rule set [2]. MyMeals reveals the share of each meal’s and each day’s energy that came from ultra-processed food. An umbrella review of 45 pooled analyses linked higher ultra-processed intake with higher risk of cardiovascular death, type 2 diabetes and common mental disorders [3].

Fiber and protein distribution

MyMeals tracks fiber against your daily goal. Meta-analyses of prospective studies found that people eating the most fiber had lower cardiovascular and all-cause mortality than those eating the least [4]. MyMeals tracks protein across the day, down to how many main meals reached 25 g: in a controlled study, spreading protein evenly across meals produced higher muscle protein synthesis over 24 hours than loading it into dinner [5].

Circadian alignment

MyMeals measures how long your eating day runs and your caloric midpoint, the time by which half of the day’s energy is in, and compares both with your 7- and 28-day averages. Meal timing is an active field: time-restricted eating has shown mixed results in trials, for example no added weight loss over calorie restriction alone in a 12-month trial [6], while later eating has been linked with less favorable metabolic markers in observational work. MyMeals maps your circadian pattern and grades the evidence accordingly.

Nutrient synergies and blockers

Nutrients change each other’s absorption within a meal. Calcium in the same meal lowers how much iron you absorb from plant foods [7], and vitamin C raises it. MyMeals applies published absorption models to flag the pairings that matter, meal by meal.

Health characteristics and diet quality

Food groups. Grams of vegetables, fruit and berries, whole grains, pulses, nuts and seeds, fish, red and processed meat, dairy and more, as the Nordic Nutrition Recommendations 2023 define them. MyMeals counts potatoes and juice separately, as the guidelines do.

Quality notes. Protein, fiber, added sugar, processing and the number of different plants, each rated low, moderate or high against published thresholds.

Glycemic load. How much a meal raises blood glucose: each food’s AI-derived glycemic index times its available carbohydrate, adjusted for the fat, protein and fiber eaten with it. MyMeals shows it as a range, never a single exact number.

Health dimensions. Four measures of your eating pattern over days, each built from food groups and nutrients with published evidence:

• Cardiometabolic quality: the dietary factors with probable or convincing evidence for blood lipids, blood pressure and blood glucose.
• Inflammatory potential: how strongly the pattern leans toward foods associated with higher or lower inflammation markers such as CRP and IL-6. This is MyMeals’ own evidence-graded model, informed by the research behind the Dietary Inflammatory Index and similar indices.
• Gut support: fermentable fiber, plant diversity and fermented foods, against additive-heavy ultra-processed food.
• Bone health: the supply of calcium, vitamin D and protein.

Diet-quality scores. Each week, MyMeals scores your eating against an established index for your region: the Healthy Eating Index 2020 in the US, the NNR 2023 food-based score and the Baltic Sea Diet Score in the Nordics, and others, such as a Mediterranean diet screener or the MIND diet, when a protocol you follow uses them. Each score needs several tracked days, and the report states how many days and components it rests on.

Protocols. If you follow a protocol, such as time-restricted eating, high-protein, low-carb or 30 plants a week, the weekly report scores your meals against it component by component.

How MyMeals grades evidence

Every link between a food and a health measure carries an evidence grade from the World Cancer Research Fund / American Institute for Cancer Research framework (2018): convincing, probable, limited but suggestive, or limited with no conclusion. How far a food moves a measure depends on both the grade and the size of the effect in the evidence.

By default, MyMeals shows only what rests on convincing or probable evidence. Exploratory measures, such as oxidative balance or a MIND-style brain-health pattern, appear only when you switch them on, and they carry the exploratory label.

Meal names and the day

MyMeals names breakfast, lunch, dinner and snacks from your own routine, not fixed clock times, and calibrates to your pattern as you go. Rename a meal, and choose whether MyMeals calibrates to the new name or treats it as a one-off.

Sovereign memory

Cloud AI is a transient lens: it looks at your food and analyzes the meal. Your health graph lives in a local-first, encrypted SQLite engine on your iPhone, where the health engine calculates your scores, so charts and reports load instantly, even in airplane mode. The graph syncs securely to your account. MyMeals never trains models on your dinners and never sells your data. Every provider is listed in the privacy policy.

What a 2D photo cannot reveal

• Mixed dishes, sauces, fried food and sweetened drinks hide their ingredients. MyMeals flags those items; calibrate them, or add the recipe once as your own food.
• Food-composition data describes typical foods. Your brand or recipe can differ, so barcodes, labels and receipts take precedence.
• Health characteristics describe eating patterns. Diagnosis and treatment decisions belong with you and your clinician; see the health disclaimer.
• Allergen data never replaces the product label. Always read it.

References

1. McDonald D, Hyde E, Debelius JW, et al. American Gut: an open platform for citizen science microbiome research. mSystems. 2018;3(3):e00031-18. https://doi.org/10.1128/mSystems.00031-18
2. Monteiro CA, Cannon G, Levy RB, et al. Ultra-processed foods: what they are and how to identify them. Public Health Nutrition. 2019;22(5):936–941. https://doi.org/10.1017/S1368980018003762
3. Lane MM, Gamage E, Du S, et al. Ultra-processed food exposure and adverse health outcomes: umbrella review of epidemiological meta-analyses. BMJ. 2024;384:e077310. https://doi.org/10.1136/bmj-2023-077310
4. Reynolds A, Mann J, Cummings J, et al. Carbohydrate quality and human health: a series of systematic reviews and meta-analyses. The Lancet. 2019;393(10170):434–445. https://doi.org/10.1016/S0140-6736(18)31809-9
5. Mamerow MM, Mettler JA, English KL, et al. Dietary protein distribution positively influences 24-h muscle protein synthesis in healthy adults. Journal of Nutrition. 2014;144(6):876–880. https://doi.org/10.3945/jn.113.185280
6. Liu D, Huang Y, Huang C, et al. Calorie restriction with or without time-restricted eating in weight loss. New England Journal of Medicine. 2022;386(16):1495–1504. https://doi.org/10.1056/NEJMoa2114833
7. Hallberg L, Brune M, Erlandsson M, et al. Calcium: effect of different amounts on nonheme- and heme-iron absorption in humans. American Journal of Clinical Nutrition. 1991;53(1):112–119. https://doi.org/10.1093/ajcn/53.1.112