Sovria from the inside

How Sovria thinks, just as she is.

For people who work in health and technology: the path of a question, the sources, the safety locks and what runs where. No promises we do not keep.

1.66 Mpassages from guidelines, books and studies

380interaction rules, each one with its source

Hereevery answer is generated on our own server, with a GPU, in Brazil

5products on the same path: the 4 apps and Artemis Nutri

The path of a question

From what you write to an answer with sources.

Safety firstEmergency signs are recognized before anything else: the answer goes out without waiting for the thinking and starts with SAMU 192, in Brazil. In emotional distress, CVV 188 comes with it.
The searchThe question becomes a search across 1.66 million passages from guidelines, reference books and studies.
The source judgeA separate judge reads the passages found and lets through only those that answer that question. A passage that looks similar but is off topic stays out.
The fixed rulesWhat cannot depend on the AI goes in as a written rule: 380 interactions between medications, supplements and foods, and the dietary pattern for each condition, each with the label or guideline it came from.
The thinkingWith the sources on the table, the model thinks before writing. That is why an answer takes about a minute.
The checkBefore it goes out, the answer passes through the safety locks: medication doses only for professionals and no promise of a cure. Where the risk is high, such as very low blood sugar, the text is fixed and has been checked.
The answer, with sourcesEvery important statement carries the number of its source, and the list comes at the end.

Our own infrastructure

Part	Where it runs	What it is
The language model	Our own server, with a GPU, in Brazil	Generates every answer here, with Sovria's method, sources and safety locks.
Vision	Our own server	The same model reads photos of meals, labels and lab tests. The SOS Restaurante photo, in Noor, is not stored.
Voice	Our own server	The apps' speech and listening run here.
The knowledge base	Our own server	1.66 million passages, each with its source.

Health data is sensitive data under Brazil's LGPD. The privacy policy (Portuguese) says what is stored, where and for how long.
Every fixed rule was checked against its source before going in.

We removed modules that promised personal predictions ("what if?") and odds: a number from a study does not become a promise to a person.
Honest follow-up is something else: try, measure how it went and adjust.

For health companies

To understand how she works inside, integrate something similar or propose a partnership, write to contato@sovriaai.com.