Our technology

Built to be checkable, not clever

Health software earns trust by being inspectable. This is how a report becomes an explanation, why each step is built the way it is, and what we deliberately refuse to do.

The pipeline

From a photograph to an explanation

Four steps. Each one can refuse to proceed, and refusing is always cheaper than being wrong.

Extraction

A photographed page becomes structured values — the test name, your number, the unit, and critically the reference range your laboratory printed beside it. This step gets the strongest model we have available, every time, on every plan. Misreading a value or a printed range is the one failure that causes real harm, so it is never the place to economise.

Matching

Extracted names are matched to a canonical list of measurands with their known aliases — the same value is printed a dozen different ways across laboratories. Anything that cannot be matched confidently is shown as unmatched rather than guessed at.

Explanation

The explanation is generated against the extracted values and sourced reference content. It is constrained to what the report says and what our sources say. It cannot introduce a threshold that is not in the sources, because there is nowhere for one to come from.

Review

Reference content is human-reviewed before it is ever shown. Every AI answer in the app carries a control to report it, and those reports are read — they are our earliest signal that something is drifting.

Deliberate choices

Why it is built this way

The model is chosen per task, not per plan

Reading a report and checking a meal are high-stakes extraction, so they get the strongest model regardless of what you pay. Conversation is where the paid tier gets a better model — because a conversation explains values that were already extracted correctly. Quality is the paid advantage; accuracy never is.

We are building our own health model

Trained on how values and printed ranges are read and interpreted — not on who anyone is. The training data carries no user identity, and it never will. Owning the model that does the reading is what turns this from an application into infrastructure.

Content is cached, not called live

Reference content is fetched, reviewed, and served from our own systems on a refresh cycle. A source going down does not take the app down, and no third party sees what you looked up.

98 languages, including spoken input

Transcription is priced per minute of audio rather than per word, so a question asked in Urdu costs what the same question costs in English. That is why speaking is affordable enough to give away to people who cannot type their own script comfortably.

We do not publish the names of the model providers or infrastructure vendors we use. They are chosen on merit and they change. If you need that detail for a compliance review, ask us and we will provide it under the appropriate agreement.

Security and data

Where your results live

Health data belongs to a person, not to a login — enforced in the database schema itself, not only in the interface.
Row-level security on every table, so a query can only ever return your own rows.
Reports are held in private storage. There is no public URL to guess.
Provider credentials are server-side secrets. They are never in the app binary and never sent to a device.
Data is stored on encrypted servers in the European Union.
Delete your account and everything in it goes — from the app, or from this website without installing anything.

Read the privacy policy Delete your account