You have a contract with client names and addresses. A colleague needs a copy for review. An AI tool could summarize it in seconds. But both paths require sending the document somewhere, and once it's sent, those names exist in places you don't control.

The usual options: spend 30 minutes manually blacking out text in a PDF editor, or send it and hope for the best. Most people pick the second one.

RedMatiq is a Mac app that sits between those two options. Drop in a document, and it scans for sensitive information using a multilingual BERT model running locally on your machine. Names, addresses, phone numbers, financial details, ID numbers. It replaces what it finds with consistent placeholders: "John Smith" becomes <PERSON_1>, "Zurich" becomes <LOCATION_1>. The same entity gets the same placeholder everywhere, across every document in a project.

Then you review. A sidebar shows every detection with the original value and its replacement. Toggle items off if they're false positives. Add manual redactions for things the model missed, like internal project names or domain-specific terms. When you're satisfied, export a clean PDF or Markdown file, or pipe the redacted content directly into an AI chat.

What it does

The core workflow is three steps:

  1. Import. Drop a PDF, Word document, Excel spreadsheet, or text file. RedMatiq parses the content, strips hidden metadata (author names, edit history, embedded comments), and runs entity detection.

  2. Review. The Privacy Inspector lists every detected entity. Three view modes let you see the original text, the pseudonymized version, and the fully anonymized output side by side. You decide what stays and what goes.

  3. Export or analyze. Generate a sanitized PDF or Markdown file. Or use the built-in AI chat, which automatically pseudonymizes your document before sending it to the model and restores real names in the response. You read a natural conversation; the AI never saw the original data.

For multi-document projects, dossiers maintain consistent entity mappings across files. If a person appears in five documents, the same placeholder is used in all five.

What makes it different

Everything runs locally. Detection, parsing, and redaction happen on your Mac. There are no cloud calls during processing. The only time data leaves your machine is when you explicitly export a file or use the AI chat, and by that point, the sensitive content has already been replaced.

Pseudonymization, not deletion. Traditional redaction blacks out text and destroys context. RedMatiq replaces values with typed placeholders that preserve document structure. AI tools can still reason about relationships between entities without knowing who those entities are.

Glass box, not black box. You see exactly what was detected, what was replaced, and what the output looks like before anything leaves your device. A "Show Sent Data" button on every chat message lets you verify exactly what reached the AI.

It's an assistant, not an automator. Automated detection gets most of the way there. The last stretch is where mistakes happen: is "Paris" a city or a person? Is "Jordan" a country or a name? RedMatiq flags what it finds and lets you make the call. We don't promise 100% detection because in privacy work, overconfidence is the actual risk.

Standard Redactions let you define your own rules. Add keywords, phrases, or internal terms that should always be redacted. RedMatiq applies them automatically across every document in a dossier, catching domain-specific content the AI model can't know about.

What it doesn't do

RedMatiq is macOS only. There is no Windows or Linux version.

Detection is not perfect. The BERT model handles multiple languages and most common entity types well, but edge cases exist. That's why human review is a core part of the workflow, not an afterthought.

It does not replace legal counsel or formal compliance processes. It's a tool that makes careful document handling practical for daily work.

The technical stack

The frontend is native SwiftUI. Detection runs on a bundled Python backend using a multilingual BERT model (Davlan/bert-base-multilingual-cased-ner-hrl) via Presidio and spaCy. Document parsing uses Docling. The AI chat connects to cloud LLMs, but only sends pseudonymized content. An MCP server lets you connect RedMatiq to Claude Desktop, Cursor, or any MCP-compatible tool with the same privacy guarantees.

Made in Switzerland.

Try it

RedMatiq is currently in public beta. Free while in beta — one-time purchase afterward. Download at redmatiq.app.

Related reading

Try RedMatiq

Local document redaction for Mac. Free while in beta.