Vera X-Ray · Client-local document privacy

Development preview

Keep the file on your device. Send only what you approved.

Vera X-Ray is IamVERA's client-local privacy layer. It detects sensitive information before AI processing, lets you review exactly what should be protected, and replaces approved values with document-bound tokens. Vera works on the context you release; X-Ray restores authorised values after the answer returns.

Client-local processing · PDF & OCR · Dutch & English · Optional local AI · Human review before release

Original — device

Dear Mrs Jansen,Your account NL91 ABNA 0417 1643 00is referenced in case 05/123456-31.

↓ X-Ray · Approved context

Dear ⟦XRAY:E1:a71c…⟧,Your account ⟦XRAY:E2:19bd…⟧is referenced in case ⟦XRAY:E3:f821…⟧.

↓ Vera · Restored after response

Mrs JansenNL91 ABNA 0417 1643 0005/123456-31

What is Vera X-Ray?

Vera X-Ray is a client-local privacy layer for professional AI workflows. It analyses documents on the client, detects sensitive values, lets the user review the findings, and replaces approved values with opaque tokens or removes supported secrets before the context enters Vera. After Vera responds, authorised values can be restored using the protected local X-Ray mapping.

X-Ray reduces unnecessary disclosure. It does not guarantee that every sensitive or indirectly identifying detail will be detected, and it does not make an AI answer true. Human review remains part of the workflow.

A confidential document should not become raw AI input

A legal file, medical report, board document or internal investigation often contains far more identifying information than an AI task requires. Uploading the complete original means the privacy decision happens after the file has already crossed the boundary.

Vera X-Ray moves that decision forward.

What does AI actually need to see to perform this task?

Detect locally

Structured identifiers, names, addresses, organisations, references and supported secrets are detected on the client.

Review before release

X-Ray shows what it found. You can mark additional text, adjust permitted replacements and inspect the prepared context.

Minimise disclosure

Reversible values become opaque references. Supported credentials and secrets are removed instead of being placed in a reversible mapping.

From confidential document to AI-ready context

  1. Open the document locally

    Choose a PDF, TXT or Markdown document in Vera X-Ray. Text-based PDFs are read locally. Scanned pages without a usable text layer can be processed with local OCR.

  2. Detect sensitive information

    X-Ray runs deterministic rules and structural validators first. It can detect supported identifiers, contact details, names, addresses, organisations, references, financial identifiers and secrets.

    An optional local recognition model can add person, organisation and city findings. The model cannot suppress deterministic findings or declare a document safe.

  3. Review the findings

    You see the proposed replacements before anything is released. Mark additional text where necessary, keep a value when disclosure is intentional, and pay additional attention to OCR-derived text.

  4. Dehydrate the approved context

    Protected reversible values are replaced with opaque document-bound X-Ray tokens. Supported secrets are removed irreversibly.

  5. Approve what Vera receives

    Delivery remains blocked until you confirm that you reviewed the prepared text. Only the approved version can be added to Vera through the X-Ray workflow.

  6. Restore after Vera responds

    Vera processes the approved context through its verification workflow. When a returned answer contains authorised X-Ray references, the corresponding values can be reconstructed on-device and shown in the X-Ray view.

Restoration authenticates the mapping. It does not certify that Vera's statement is factually correct.

Useful context without the original identity

Original — client side

Dear Maria Jansen,The payment should be made toNL91 ABNA 0417 1643 00.

Approved for Vera

Dear ⟦XRAY:E1:a71c…⟧,The payment should be made to⟦XRAY:E2:19bd…⟧.

Restored after response

Maria JansenNL91 ABNA 0417 1643 00

The external AI works on the approved context and document-local references instead of the protected values themselves.

X-Ray does not invent a second identity

Traditional pseudonymisation can replace one real identity with another realistic-looking identity. X-Ray uses opaque random references instead.

Maria Jansen⟦XRAY:E1:6f10c2…⟧

Entity continuity without exposing the value

Repeated values can share a document-local entity reference such as E1.

Random token per occurrence

Each occurrence still gets its own 128-bit random component.

No cross-document identifier

Entity numbering starts again for every document.

Secrets expose less

Removed passwords, PINs, keys and supported credentials receive no reusable entity reference and are not retained as reversible originals.

An entity reference indicates that X-Ray considers occurrences equivalent within that document. It is not a global identity and does not prove that two people with the same name are the same person.

Local AI only where it adds value

Vera X-Ray does not send a confidential document to a generative model and ask whether the document is private.

The first detection layer is deterministic: checksums, known structures, document cues and language-specific rules handle information that does not require AI.

For harder free-text entities, the user can optionally install a locally running recognition model. The currently enabled model categories are person, organisation and city. Its findings are additive and remain subject to review.

Rules and validatorsOptional local NERHuman review

Model files are downloaded separately and verified against pinned SHA-256 digests. Document text is not sent to the model host for inference.

The private mapping stays protected on the client

Encrypted X-Ray mapping

Reversible values and mapping metadata are stored in an authenticated encrypted X-Ray envelope.

Random tokenisation

Protected values use opaque random references, not hashes of the original value.

PIN-protected saved mappings

Approved mappings that must survive the current panel session can be protected locally with a PIN.

Automatic lock

Saved mappings lock after inactivity. There is no silent server recovery path for a forgotten local PIN.

The original document and private X-Ray mapping remain client-side during document preparation. External AI providers receive the context you approved. Restored values may subsequently be displayed in Vera when you explicitly request restoration.

Text PDFs and scanned pages

PDF with text layer

Text is extracted locally before detection.

Scanned PDF

Pages without usable text can be rendered and processed with local OCR.

TXT & Markdown

Plain text can be processed directly.

OCR can introduce recognition errors. OCR-derived pages therefore require additional human attention and do not become “safe” simply because no detector fired.

X-Ray controls what AI sees. Vera verifies what AI says.

Input assurance

Vera X-Ray

What is AI allowed to see?

  • detect sensitive values;
  • minimise unnecessary disclosure;
  • review the outbound context;
  • pseudonymise or remove;
  • approve before delivery.
Output assurance

Vera

What should I trust in the answer?

  • generate with selected models;
  • fact-check claims;
  • challenge reasoning;
  • verify available sources;
  • expose disagreement;
  • keep a traceable verification record.

Privacy before processing. Verification before action.

Built for professionals who cannot simply paste the file into AI

Lawyers & legal teams

Prepare correspondence, contracts and selected case material for AI-assisted analysis while reducing unnecessary disclosure of client identifiers.

Notaries

Protect names, addresses, account details and document references while retaining the context needed for document work.

Healthcare and occupational health

Prepare text for editorial or analytical support while reducing exposure of direct identifiers. X-Ray is not a clinical decision system and does not claim complete medical-context detection.

Compliance & security

Remove supported credentials and prepare investigations, policies and internal documentation before AI processing.

Boards & executives

Review exactly what confidential business context will enter an AI workflow before releasing it.

Ordinary AI upload vs Vera X-Ray

Ordinary AI upload vs Vera X-Ray
Ordinary AI workflowVera X-Ray
Upload the original filePrepare the document client-side first
Privacy decision happens after uploadPrivacy decision happens before release
Little visibility into what leftExact outbound text is reviewable
Original identifiers may enter the promptProtected values become document-bound tokens
Credentials can travel with the fileSupported secrets are removed irreversibly
No local mappingProtected encrypted X-Ray mapping
AI answer contains placeholdersAuthorised values can be restored after the response
“No finding” may look safeUser review remains mandatory

Inspect what X-Ray actually demonstrates

We do not use “privacy-preserving” as a substitute for evidence. X-Ray is tested with deterministic regression suites, synthetic document corpora, model parity checks and end-to-end browser flows.

Those tests describe the cases that were tested. They do not prove that every future document will be detected correctly.

Deterministic regression tests

Checksums, identifiers, correspondence layouts and previously reproduced failure cases.

Synthetic evaluation corpus

Dutch and English document fixtures with explicit gold spans.

Local-model parity

The browser implementation is tested against the pinned reference model and export pipeline.

End-to-end browser flow

PDF, OCR, review, approval, delivery and restoration are exercised through the built extension.

Explore Vera's evidence approach

What Vera X-Ray does not claim

  • Detection is not perfect.
  • Indirectly identifying combinations of ordinary facts can remain.
  • OCR can make mistakes.
  • Local model findings are advisory.
  • A user can deliberately keep a detected value.
  • The current browser product is still in development.
  • Firefox support is required but is not the current production target.
  • X-Ray does not verify that an AI-generated statement is true.
  • X-Ray does not replace organisational access control, device security or professional judgement.
  • Restored values are sensitive again once displayed or exported.

X-Ray reduces unnecessary disclosure. It does not turn confidential work into risk-free AI use.

Questions about Vera X-Ray

What is Vera X-Ray?

Vera X-Ray is IamVERA's client-local document privacy layer. It detects sensitive information before AI processing, lets the user review what should be protected and replaces selected values with opaque document-bound tokens before the approved context enters Vera.

Does Vera X-Ray anonymise a document?

Not in the strict sense. Reversible replacement is pseudonymisation because authorised values can later be restored from a protected local mapping. Some supported secrets are removed irreversibly. Unique combinations of remaining facts may still identify a person.

Does the original document go to an AI model for X-Ray detection?

The X-Ray document-preparation flow runs locally using deterministic rules, local PDF/OCR processing and, when enabled, a local recognition model. The later Vera verification step is separate and operates on the context you approved.

What reaches external AI providers?

The approved context used in your Vera session may be sent to the AI providers selected for that workflow. Values protected by X-Ray are represented by X-Ray tokens unless you deliberately chose to retain them.

Can Vera X-Ray process scanned PDFs?

Yes. Pages without a usable text layer can be processed with local OCR. OCR errors remain possible, which is why the resulting text still requires review.

Does the local recognition model receive my document over the internet?

No document text is sent to the X-Ray model host for inference. When the optional model is installed, pinned model files are downloaded and inference runs on the client.

Can Vera X-Ray restore Vera's answer?

Known authorised tokens can be restored while the corresponding local mapping remains available and unlocked. Restoration verifies the mapping, not the factual correctness of the generated answer.

Can I download Vera X-Ray now?

Vera X-Ray is currently in development and is not yet offered as a general public download. Contact us to discuss a demonstration or early deployment workflow.

Start with what you do not want to share.

See how client-local document preparation can fit into your Vera workflow — before confidential context reaches external AI.

X-Ray controls what AI sees. Vera verifies what AI says.