A Zyric product
In DevelopmentTessika
A lab report photographed on a phone, a scanned invoice, a passport, a handwritten note. Tessika reads it the way a person would, masks the personal data on it, and returns one clean record in the shape your system expects.
01Overview
Tessika is a document engine for data that arrives in every shape. It reads straight from the image rather than through OCR, so a crooked phone photo or a faded scan still comes out right, and it takes PDFs, images including iPhone photos, Word and Excel files, and plain text.
Before anything is extracted, the personal data on the page is found and masked. What comes out is one canonical record per fact, each with a confidence and a trail back to where it came from, ready for the system that builds on it.
02Process
Step 1: Read
Every page is read as an image by a vision-language model, so tables, stamps, handwriting and bad scans are understood the way a person sees them. Long documents are split into batches and merged back, page numbers intact.
Step 2: Protect
Names, dates of birth, addresses, ID and account numbers are located and covered. A strict setting also covers faces, signatures and the machine-readable zone of an ID or passport, and leaves what must stay visible, such as an expiry date.
Step 3: Verify
The masked page is read again, and anything personal that is still legible is covered too. The result reports whether the re-read came back clean, rather than assuming it did.
Step 4: Structure
Each document type maps to a canonical record. Health, legal, finance, insurance and identity ship today, and you can hand in your own output schema, or describe what you want in plain language.
03By design
Runs where your data lives. Tessika runs in your own environment, per region if needed, on an open-source model that can be swapped without touching the pipeline. No document has to go to a third-party AI service.
Originals are not the point. The engine keeps a reference to the original, never its bytes, and stores a masked page only when you ask for one.
Audited, not assumed. It ships with its own audit: documents with known planted personal data, and a leak check that reads the masked output back independently of the model. A page that only looks redacted fails it.
04Honest status
In development and working end to end. Masking is not yet perfect on every hard scan, which is exactly what the built-in audit exists to catch. Available on request, adapted to your documents and your infrastructure.
Your documents
Tell me what kind of documents you deal with and what you need out of them, and I will tell you honestly whether Tessika fits.
Or write directly to chris@zyric.de.