Document Scanning
Reads scanned PDFs and images on your own hardware, fully offline, and feeds the text into your local knowledge store.
A lot of what an organisation actually needs to work with isn't born digital. It's a scanned contract, a photographed form, an old case file that only exists as an image of a page. Most AI tools either can't read it or send it off to be read somewhere else. Mickai's OCR engine reads it on the machine it's running on, and nowhere else.
It runs local
The engine reads document images entirely on the hardware it's installed on. No document, page or image is sent anywhere to be read. It needs a capable NVIDIA GPU to run today; there is no cloud fallback.
It knows what it's not
Every result the engine produces is labelled as a transcription of a scan, not text extracted from the document itself. That label travels with the output. Verify against the original before you quote it.
It fills a real gap
When a PDF has no real text layer, most systems either fail silently or quietly guess. Mickai's document pipeline says so explicitly and hands the page to the OCR engine, so a scanned document stops being a dead end.
What it reads
Scanned PDFs, and common image formats including PNG, JPEG, BMP, WEBP, TIFF and GIF. When a PDF arrives with no genuine text layer, the document pipeline says so rather than guessing, and hands the page to the OCR engine instead of quietly skipping it. The text that comes back is written into your own local knowledge store, with an entry in the audit trail recording that it happened.
Where it stands
The OCR engine works today on a validated configuration: a workstation with a capable NVIDIA GPU. There is no CPU-only path yet. Accuracy has been measured and confirmed on Latin-script pages; we haven't measured other scripts, so we're not claiming a result for them. Running it on a machine we haven't already set up and tested is still open work, not a finished, ship-anywhere feature. We're not publishing speed or throughput figures because we don't have ones we'd stand behind as general claims, only single-page measurements on one test machine.