Records
Every outgoing message gets a number, a timestamp and a seal. The log records what went out and when.
Schwerpunkt: Lokale KI und Pseudomizer. Zusammen. Mehr
Branchenpartner gesucht. Partner werden
Ihre Branche: Rechtsanwalt oder NotarSteuerberaterArzt, Zahnarzt oder PraxisBiogas und EnergieIndustrie und MaschinenbauSchule und BildungLandwirtschaftHandwerkBüro, Verwaltung, Kommune
Focus · MLKI and Eigenkern Pseudomizer
The local AI MLKI stays in your house. The Pseudomizer creates a shadow copy of your documents in which names and confidential details are replaced. Only after a check chain may this copy go to a large cloud AI. The leak rate at the exit is 0.00 %: guards detect, the judge stage blocks when in doubt, healing cleans up.

Ms MusterAlpha of Muster LtdBeta Ltd asks for a call back about invoice RE-2026-0815Identifier 1 for €4,800.
Example with invented details.
The leak rate
The leak rate is the share of sensitive details (names, companies, identifiers) that still appear in full in the shadow copy after pseudonymisation. At the exit it is 0.00 %, because guards, judge stage and healing work together.
The figures are technical measurements with test data, not a legal certificate and not a promise for your documents. We measure which rate is reached with your documents in the test with you.

The shadow copy carries the meaning of your document, but not your names and identifiers.
Briefly explained
Names are replaced by placeholders, the key stays with you.
Example with invented values
The path of a document
Example with invented values
Hybrid
Originals stay in the house. The cloud AI only sees the twin RAG of the shadow copies.
Example with invented values
The data recorder
Like the flight recorder in an aircraft: every message to a cloud AI is stored numbered and sealed. You can look up what went out, and prove it.
Example with invented values
Every outgoing message gets a number, a timestamp and a seal. The log records what went out and when.
The entries are linked as a seal chain. Anyone who changes or deletes an entry afterwards breaks the chain, and it shows.
Checking is done by the check chain before it. The data recorder sits at the exit, sees the real data and cannot be lied to by the services behind it.
Status: connected in our test environment; connection to all routes to cloud AI services is in progress. The log shows what went out. It does not replace a legal certificate.
The check chain
Each step is a separate, independent check. If anything is found in a step, nothing goes out.
Your documents are on the device in your business. MLKI reads and sorts them there.
Several detection modules run in series and search for names, companies, addresses and identifiers. A second, independent detection catches what the first one misses.
Details found are replaced by related words so that the meaning of the document is preserved. Figures, amounts and quantities stay unchanged.
An independent check chain searches every job once more for remnants, using dictionaries and rules per subject area. Then the judge, a separate AI stage, rules on every suspicious passage in the context of the sentence. If it is not sure, or the judges disagree, the passage is blocked. If anything is found, nothing is sent.
Only the checked shadow copy goes to the chosen cloud AI. It never sees the original.
The answer comes back. Only at your site are the real details put back in.

An independent check chain inspects every job. If anything is found, nothing goes out.
Security and documentsSubject-specific check chains
The check chain works per document type and subject area with its own dictionaries and rules. What is sensitive in a tax practice differs from what is sensitive in a medical practice or a biogas plant.
Clients, case numbers, bank details and addresses.
To the industryNames, addresses and details that fall under professional secrecy.
To the industryMetering points, plant identifiers, contract partners and bank details.
To the industryCustomer, supplier and project names as well as case numbers.
To the industryWe set up further subject areas together with industry partners. The dictionary of a subject area is built from real documents of that industry.
Stated openly
No method recognises every detail. That is why we check twice and independently, block on every finding and measure with test data.
Our measurements are technical evidence, not a legal certificate. We clarify legal bases and access rights with you before the start.
The Pseudomizer is under development. German, English, French and Nordic languages are tested. Cyrillic and East Asian scripts need their own models and have not yet been tested.

The return path with the real details runs at your site, not in the cloud.
One conversation, one real task from your daily work. We use it to show what AI and automation can do for you.