AI Guard — Thai PII Redaction

Mask Thai personal data (PII) before sending it to an external AI, then restore the real values locally. Everything runs on your machine — raw PII never leaves the device (PDPA-friendly).

PSU Future Tech Challenge 2026 · AI Innovation for Future Society (Prototype)

What it does

• AI Guard — on ChatGPT, Claude, Gemini, Grok, Perplexity, or GLM / Z.ai, replace PII with tokens ([ชื่อ_1]) or realistic fake data before sending, then restore the real values from the reply. (On any other page, the docked side panel does the same by paste.)
• True PDF redaction — black out PII in a PDF (removed from the text layer), returns the redacted file + before/after preview.
• PDPA report — risk score plus Section 26 sensitive-data flags (health, religion, …).

Detection runs locally: regex + checksum (Thai ID mod-11, phone, email, …) and Thai NER (PyThaiNLP). No cloud AI is used to detect.

Installation

The recommended way to run AI Guard is the desktop app: a native installer that bundles the backend, so there is no Python setup. A from-source / pip path is also available for developers.

Option A · Desktop app installer (recommended)

1. Download the installer for your platform from the Releases page:
   • Windows (tested, primary platform): AI.Guard_<version>_x64-setup.exe
   • macOS (experimental — built in CI, not yet verified on real hardware): AI.Guard_<version>_aarch64.dmg
   • Linux (experimental — built in CI, not yet verified on real hardware): AI.Guard_<version>_amd64.deb or AI.Guard_<version>_amd64.AppImage
2. Run the installer and launch AI Guard. The backend is bundled and starts with the app — no Python, no separate install, works offline.
3. The app is not code-signed. On first run, Windows SmartScreen may show an "unknown publisher" warning — click More info → Run anyway to continue.

Option B · From source (developer alternative)

Prerequisites: Python 3.11+ and git.

git clone https://github.com/Teerapat-Vatpitak/thai-pii-redaction.git
cd thai-pii-redaction

./run.ps1     # Windows (PowerShell)

./run.sh      # Linux / macOS / git-bash

The script creates a virtual environment and installs dependencies on first run (a few minutes). The first time Thai NER runs it downloads a ~2 MB model (needs internet once). This starts the same backend the desktop app bundles, at http://localhost:8000 — useful for the browser extension UI below, or for hitting the API directly.

Verify either option: open http://localhost:8000/api/health → you should see {"status":"ok"} with the current version.

Optional: browser extension (ChatGPT / Claude in-page UI)

The extension talks to the same local backend, so it works with either install option above (desktop app or from-source).

1. Open chrome://extensions in Chrome / Edge (any Chromium browser).
2. Turn on Developer mode (top-right).
3. Click Load unpacked and select the extension/ folder from this repo (clone or download the repo to get it).
4. Pin the AI Guard extension so its toolbar icon is handy — clicking it opens the docked side panel. The in-page bar activates on chatgpt.com, claude.ai, gemini.google.com, grok.com, perplexity.ai, and chat.z.ai / chatglm.cn.

See extension/README.md for details.

Using it:

1. Type a prompt containing PII into the chat box.
2. Click Mask PII (the floating AI Guard bar) — your text becomes tokens or fake data.
3. Send it with the site's normal Send button.
4. When the AI replies, click Restore PII to see the real values.

Troubleshooting

• "Backend offline" in the extension — the backend isn't running. Launch the desktop app, or start it via the run script (Option B).
• Port 8000 already in use — close whatever is using it, or stop a previous backend instance.
• SmartScreen blocks the installer — expected for an unsigned build; choose More info → Run anyway.
• Extension bar doesn't appear — reload the ChatGPT/Claude tab after loading the extension.

Mask modes

Mode	Output	Use when
token (default)	[ชื่อ_1], [โทรศัพท์_1]	you want masking to be obvious
surrogate	realistic fake data (valid checksums)	you want the AI to read it naturally

Switch in the extension side panel.

Try the examples

In examples/ (all PII is fabricated): three realistic Thai prompts and a sample Thai PDF. Explore the API at http://localhost:8000/docs, or:

curl -X POST http://localhost:8000/api/sanitize \
  -H "Content-Type: application/json" \
  -d '{"text":"ผมชื่อสมชาย ใจดี โทร 081-234-5678","mode":"surrogate"}'

Optional: semantic sensitive detector

Catches free-form Section 26 content keywords miss (e.g. "ป่วยเป็นเบาหวาน") via a MiniLM model:

pip install -r requirements-ml.txt   # large; the feature self-disables if absent

(Not included in the desktop app or the packaged builds.)

Privacy

The token↔original vault is in memory only — never written to disk or sent over the network. The extension stores only a session_id. The external AI sees only masked text.

Tests

PYTHONUTF8=1 python -m pytest

More

Architecture and module map: CLAUDE.md.

License

Apache License 2.0 — see LICENSE and NOTICE. PDF handling uses the permissively licensed pypdfium2 / reportlab / pdfplumber (PyMuPDF/AGPL is no longer used).

Build the desktop app yourself

See desktop/README.md and packaging/README.md for the Tauri build (bundles the backend via desktop/build-sidecar.ps1).