Selected work
Tansees
AI-powered text extraction that converts images and documents into editable text, with support for Arabic documents.
The challenge
Arabic breaks OCR in ways English never does. The diacritics go first, those small marks above and below the letters that most engines read as noise and throw away, which is fine until the document is a Quranic text, a line of poetry, or a school curriculum where a single fatha changes the meaning. Then mixed text: put an English number inside an Arabic sentence and watch the line come back reversed. Tables collapse into a pile of digits with no rows. And handwriting is a disaster, notes and old manuscripts that global tools simply give up on. None of this is a bug. It's what happens when Arabic is handled as a translation layer over an English-first engine.
What I did
Built Tansees on Python and Flask with Mistral's vision model doing the reading, and made one decision early that shaped everything else: no single engine wins on every document, so stop pretending otherwise and let the user choose. That became three levels. Fast costs one credit and handles clean printed books and reports. Advanced costs two and runs a vision-language model that reads the Arabic sentence in context instead of letter by letter, which is what keeps diacritics in place. Super costs four and exists for the cases that defeat everything else, handwriting, complex tables, multi-column layouts. Nobody pays for heavy processing on a simple invoice. Then I fought the small frictions. Paste a screenshot with Ctrl+V, no saving the file first. Export to Word, Markdown, or plain text. Use it with no account at all, since asking a lawyer to register before scanning a contract is a strange thing to do. Files get deleted from the server the moment extraction finishes, no archive, no backups, and I put that on the landing page instead of burying it in a policy. I also refused to publish a benchmark table claiming we beat named competitors. The comparison on the site describes typical behaviour of tool categories on Arabic documents and says so out loud. Upload your worst file and judge it yourself, that's the honest version.
The result
Three extraction levels, four export formats, and files deleted from the server the second the text comes out. A visitor with no account gets a free extraction every day. A free Google account gets four credits daily and unlocks the Advanced level. Pro is 9.99 SAR a month for 500 monthly credits with no daily ceiling, and there's an unlimited tier at 99.99 SAR, both with a five day trial before anything is charged. Source is public on GitHub.
