Data and AI

Three jobs, and only three:

• Builds your week. It picks 21 meals from a catalogue we maintain, arranged around the symptoms you picked, the foods you said you like, the foods you said you never want to see, and any allergens you flagged.
• Scores each meal for you. The number is how well that meal lines up with what you told it is bugging you. The same meal gets a different score for a different person, because it is being measured against a different question.
• Writes the one line under the score. Every score carries its reason. That rule is enforced in the app's own code, not left to the model.

Meal photographs are generated separately, from the name of the dish alone, for example the words “overnight oats and berries”. Nothing about you is part of an image request, and each picture is generated once and reused for everyone who gets that meal.

• It does not diagnose anything, and it is not making a clinical judgement about you.
• It does not act on its own. It generates a plan; it does not buy anything, message anyone, or change your settings.
• It does not see photographs of you or of your food. Nuvia has no camera access at all.
• It does not read anything you have not given it. There is no access to your contacts, your location, your health app, or any other app on your phone.
• It does not set the price of anything, or decide what you are charged.

Three sources feed a plan, and they are worth separating:

1. What you told us

The symptoms you picked and the follow-up detail on each, your goal, your date of birth, gender if you gave it, activity level, how you eat, foods you love, foods you skip, allergens, and your daily check-ins. This is the part that makes the plan yours, and it is the only part that is about you.

2. Our meal catalogue

The dishes, their ingredients and their preparation are a catalogue we author and curate. The model chooses from it and scores against it; it is not inventing food from nothing. When we add to the catalogue, a person reviews what goes in.

3. General nutrition knowledge inside the model

The reasoning about which foods are commonly associated with which symptoms comes from the model's general training, which is broad public knowledge rather than a specific clinical source we control or can cite. This is the weakest of the three sources, and it is the reason the app's language is hedged: “linked to”, “may support”, “commonly associated with”. Those words are doing real work. They are not marketing softeners.

We do not cite studies, and you should not read a score as evidence. A score is a planning tool that reflects general association, not a finding about your body.

• Anthropic (Claude): generating your weekly plan, scoring meals, and writing the reason lines.
• fal.ai: generating the meal photographs, from dish names only.

What is sent

When a plan is generated or a meal is scored, the relevant part of your profile goes to Anthropic: the symptoms you picked, your goal, your allergens, how you eat, the foods you love and the foods you skip.

What is never sent

Your name, your email address, your account identifier, your payment details and your device identifiers are not part of those requests. The model receives a description of preferences and symptoms, not an identified person.

Training

Under our agreement with Anthropic, the content of our requests is not used to train their models. We do not sell your data, and we do not license it to anyone to train anything.

The model is not left alone with the output. Before a plan reaches you it is validated against your allergens and your stated eating style, and a plan that fails that check is regenerated rather than shown. The meal catalogue is curated by us, and additions are reviewed by a person before they can appear in anyone's week.

If a plan still comes out wrong, that is a bug and we want to hear about it. Write to support@nuvia.app with the day it is on and what looks wrong.

Known and real, not hypothetical:

• Generated text can be wrong. Ingredient lists, quantities and preparation steps are produced by software and are not a substitute for the label on the food you actually buy.
• Allergen handling is a filter, not a guarantee. We plan around what you flag, but if you have a severe allergy, treat the app as a suggestion and check every label yourself.
• Scores are relative, not absolute. A 92 means “lines up well with what you picked”, not “healthy” and not “proven”.
• It only knows what you tell it. It has no access to your medical history, your medication, or your bloodwork, and it cannot account for any of them.
• Association is not causation. A pattern in your check-ins is a prompt to look closer, not a finding.

• Change the inputs. Everything the model is given about you is editable in the app, under You. Change it and next week's plan changes with it.
• Export it. Settings, then Export my data, gives you a machine-readable copy of what we hold.
• Delete it. Settings, then Delete account, removes your profile, plans, check-ins and preferences.
• Stop using it. Cancel the subscription in your iPhone Settings. No plan is generated for an account that is not active.

There is no way to use Nuvia without the AI, because generating your week is what the app is. If you would rather not have a model involved in planning your food, Nuvia is not the right tool.

If we change which models we use, what they are given, or what they are used for, we will update this page and change the date at the top. Questions about any of it: support@nuvia.app. The privacy policy and the terms of use cover the rest.