MenuClose
PL
← All projects

Prototype · 2026

Anonimizator

Masking personal data before text goes to an AI chat — entirely in the browser.

  • AI / LLM
Abstract dot field with concentric circles on a green background

A browser app with no backend. You paste text, review the detected personal data, copy the masked version into any AI chat, and when you paste the reply back, the original values return to their places.

Problem

People paste contracts, emails and documents full of names, national ID numbers and addresses into AI chats. Typical anonymisation tools send the text to their own server, which moves the problem instead of solving it.

Solution

A detection core written in Rust and compiled to WebAssembly, with detectors for Polish data. Nothing leaves the browser, which a canary test watching four leak channels verifies. A separate benchmark measures how much masking costs in model answer quality.

  1. Text with personal data
  2. Rust/WASM core: detection
  3. Masked text
  4. Any AI chat
  5. Originals restored
Data never leaves the browser
bytes of text sent to a server
0
leak channels under constant test
4
judgements in the quality benchmark
1,440

The canary test

A Playwright end-to-end test watches the network, the clipboard, browser storage and the DOM. So that it never gives a false sense of safety, it is regularly run against a deliberately leaking bait page — and it has to catch it.

What anonymisation costs

The benchmark generates model answers for original and masked texts (360 generations), and two independent judges grade them blind (1,440 judgements). Judge agreement is chance-corrected, so the result reflects real quality loss rather than noise.

Related projects