Atlas AI
Built for Arabic and Latin scripts and for documents from all over the world. This page says what each technology does today, and what is still in preparation.
Our technology
Each building block is part of one or more products. Anything not yet available is marked "Coming soon".
Authenticity, consistency and validity
The document is recognised (country, type, version), its fields are extracted, then checked against each other and against what the customer declared.
Arabic, French and English
Character recognition and vision-language models read the document in its own language and tie each value to the right field, with a confidence level.
MRZ checked; NFC chip coming soon
The machine-readable zone is read, its check digits verified, then cross-checked with the printed fields. Chip reading will happen in the mobile SDK.
Selfie matched to the document photo
The face is detected on the document and on the selfie, then compared: the score is set against your threshold, and the result stays explainable.
Passive, against presentation attacks
Passive detection on a single selfie: no gestures required. It flags known presentation attacks.
In preparation
Detection of generated faces and of video streams injected into the camera is in preparation. It is not available yet.
Photocopies, screen replays, MRZ mismatches
The document image is analysed to spot captures that are not the original, and the MRZ gives away edited fields.
One name, recognised in both scripts
The same name is spelled several ways in Latin letters, and differently again in Arabic. Screening brings these variants together and explains every score.
How we work
Bundled models are listed with their source, licence and checksum, and verified on every build.
Each check returns a status, a score and readable reasons, the same in the API, the console and the webhooks.
Our models are evaluated on synthetic test sets; an evaluation on real data, collected with consent, will come before any published figure.
Ambiguous cases go to review: your analysts decide, and their decisions stay logged.
See it in action
The sandbox provides synthetic identities and documents: you see every check, its score and its reasons, without any real data.