GDPR Guardian · launch video · all on-screen text, in order
Silent video (no voice, no music): this file lists every on-screen text at the moment it appears.
176.2 s · 1920 x 1080 · 30 fps

[00:00] GDPR Guardian
[00:00] Mr Jean Dupont, 06 12 34 56 78
[00:02] Mr [PERSON_1], [PHONE_1]
[00:03] Anonymise your documents, irreversibly, before they reach an AI.
[00:06] An experimental, open-source tool to help respect the GDPR.
[00:07] A filter to plug into the AI tools on the market.
[00:07] gdprguardian.com · gardienrgpd.fr
[00:15] Every day, real documents end up in AI assistants.
[00:15] Contract — Jean Dupont
[00:15] HR file — FR76 3000 6000
[00:15] Medical record — Type 2 diabetes
[00:16] E-mail — c.martin@example.com
[00:18] AI assistant
[00:20] Personal data ends up with a third party.
[00:24] Today’s answers fall short.
[00:25] Manual redaction — Slow, and names slip through.
[00:27] Reversible pseudonymisation — The mapping table is still personal data.
[00:30] [REDACTED] everywhere — The AI loses the meaning.
[00:32] Black boxes — Nothing explains the masking.
[00:35] Remove the names, keep the meaning.
[00:36] Mr Jean Dupont · jean.dupont@example.com · 06 12 34 56 78 / IBAN FR76 3000 6000 0112 3456 7890 189 / Mr Dupont will call Ms Claire Martin back. / DUPONT Jean will pay invoice no. 2026-0042.
[00:37] real output, version 1.0.1
[00:37] Random 256-bit key, for this document only
[00:42] Mr [PERSON_1] · [EMAIL_1] · [PHONE_1] / IBAN [IBAN_1] / Mr [PERSON_1] will call Ms [PERSON_2] back. / [PERSON_1] will pay invoice no. 2026-0042.
[00:42] Key destroyed: no way back.
[00:43] Same person, same tag.
[00:46] Guillaume annoys Nadège. Nadège hits Guillaume.
[00:48] Naive masking — [NAME] annoys [NAME]. [NAME] hits [NAME].
[00:49] Who hit whom?
[00:52] GDPR Guardian — [PERSON_1] annoys [PERSON_2]. [PERSON_2] hits [PERSON_1].
[00:52] The roles survive, the identities do not.
[00:57] Six principles
[00:58] Irreversible — 256-bit key destroyed, no table.
[01:00] Consistent — One person, one tag, most of the time.
[01:03] Verified — Official check digits.
[01:04] Explainable — A rule per tag, no LLM.
[01:07] Fast — 2 KB in under a millisecond, on CPU.
[01:10] Nothing kept — All in memory.
[01:13] Three ways in with a key, one engine
[01:14] Website — With your API key, drop, then download.
[01:14] MCP — Claude and other agents, with your key.
[01:14] REST API — /v1/anonymize/text /v1/anonymize/file
[01:20] GDPR Guardian — One fresh process per request, nothing kept.
[01:23] Anonymised Markdown
[01:24] Your AI
[01:25] Over MCP, send the file: it never goes through the model. Pasted text has already left.
[01:31] 100% local, no key: Claude skills (engine bundled), Docker without network.
[01:36] Copy the repository, make your own version.
[01:36] $ git clone https://github.com/FlowLineIntegration/GDPRGuardian
[01:38] Code published on GitHub.
[01:39] MIT licence — Use it, change it, share it.
[01:42] Your rules — In the call, never stored. — CLT-204518 → [CLIENT_ID_1]
[01:45] Your server — docker compose deploy/README.md
[01:47] Contribute — Improve, propose: /issues
[01:50] Tested, measured, published
[01:51] 6,600+ — automated tests, Windows and Linux
[01:53] 97.2% — of 450,000 synthetic documents with no leak
[01:56] 84% — of difficult documents with no leak: 51 ways to trap a text
[02:01] 14 — annotated real-text datasets, 6 of them French, 8,178 documents
[02:04] 95.1% — of BOAMP notices with no leak, 98.3% of mentions masked
[02:08] 97.5% — of BODACC notices with no leak, 99.5% of mentions masked
[02:13] How we grade ourselves
[02:13] Confidence index
[02:13] 96.9%
[02:14] 95% CI: 96.2–97.6%
[02:14] Mentions masked: 99.3% · precision: 99.2% BODACC, 87.7% BOAMP · neutral text masked by mistake: 0.07%
[02:20] Normalised documents: share with no readable direct identifier, protocol 1.1.
[02:20] Run of 26 October 2026: 1,019,385 documents, protocol published beforehand.
[02:27] Known limits
[02:27] Published apart: old newspapers (low-quality OCR), MAPA FR, other languages.
[02:31] Real corpora already seen while tuning: optimistic figures.
[02:34] German, Spanish, Italian, Dutch, Portuguese names: mostly missed (none without France).
[02:38] A first barrier, not a GDPR certification.
[02:41] The protocol: gdprguardian.com/#protocol
[02:44] GDPR Guardian
[02:44] An experimental, open-source tool to help respect the GDPR.
[02:48] Try it, improve it.
[02:50] gdprguardian.com · gardienrgpd.fr
[02:50] github.com/FlowLineIntegration/GDPRGuardian
[02:51] Offered by Flow Line Intégration
[02:51] Guillaume CLEMENT, CTO Data & AI
[02:51] Design: Claude Design
[02:52] MIT licence, no warranty · version 1.0.1