Clean before AI

Changelog

Every important Clean before AI release: what we improved, how the measurements changed and what still doesn't work well. We don't hide weaknesses.

1. v1.33 October 2026 · web tool + Chrome extension

   Our own local AI model, known names and a confidence light

   • Our own AI model (~53 MB, 17 languages) finds names, nicknames, addresses and companies the rules don't know. Bundled in the extension and works offline; in the web tool it downloads once when switched on and stays in your browser. Replaces the previous optional AI layer.
   • Known names: colleagues, clients and projects are always hidden — in any form, in capitals and without accents. CSV import; organisations can distribute the list through a Chrome policy.
   • Confidence light: doubtful words appear in a “Check” row — one click keeps them in the text.
   • Detection: chats (speakers, nicknames, Cyrillic), declined addresses, small businesses without a legal form, form field labels, gaps found by the blind test.

   Measurements

   Regular documents	96.7 % → 98.2 %
   Hard documents	89.8 % → 95.6 %
   Real public forms and contracts (new)	98.2 %
   Blind test — texts the engine had never seen (new)	92.9 %
   False hits on texts without personal data	0 → 0

   Known limitations

   • On unknown texts (blind test) roughly one item in 14 stays unmarked. Review the highlighted parts before sending.
   • Informal chats are the weakest area: nicknames, lowercase names, typos.
   • Some languages (English, French, Slovak) have only a few hard test documents, so their hard-document figure is not yet reliable.
2. v1.2.12 October 2026 · Chrome extension

   Faster pasting in Gemini

   • Gemini: long text (37,000 characters) pastes in 0.2 s instead of 6.2 s; the cursor stays after the text.
   • Companies written as “D.O.O” without the final dot and “D.0.0.” from scans.
   • Terms of use and a “Support the project” link in the panel.
3. v1.230 September 2026 · web tool + Chrome extension

   17 languages and real documents

   • New languages: Polish, Dutch, Romanian, Czech, Slovak, Hungarian; English for the UK and US. Language is detected per paragraph, labels follow the language of the text.
   • Extension interface in 14 languages.
   • Real documents (public contracts and forms): domestic account numbers, IBANs across line breaks, health insurance numbers, phones, forms, layout-aware PDF reading; fewer false hits.
   • “Also hide amounts” switch; the side panel opens by itself when cleaning.

   Measurements

   Test documents	136 → 263
   Regular documents	96.5 % → 96.7 %
   Hard documents	89.1 % → 89.8 %
   False hits on texts without personal data	1 → 0

4. v1.128 September 2026 · web tool + Chrome extension

   First public measurements

   • “How well it works” section: share of personal data found on test documents the engine had not seen, and false hits — re-checked after every change.
   • Spanish and Portuguese, passwords and keys, server names, references, tables.

   Measurements

   Regular documents (136)	96.5 %
   Hard documents (28)	89.1 %

5. v1.027 September 2026 · web tool + Chrome extension

   First version

   • Local cleaning of personal data in the browser: declined names, consistent labels, real data restored in the AI's answer.
   • IDs with checksums (national IDs, tax numbers, IBAN …) for Slovenia, the Balkans, Germany, Austria, Italy and France.
   • PDF and image export with redacted areas.
   • Chrome extension: ChatGPT, Claude, Gemini, Copilot, Perplexity, DeepSeek, Mistral, Grok and more; real data only in the side panel.