AI image redaction tool
Manual redaction is slow and easy to get wrong. Redacted combines OCR with a small named-entity model to find personal information for you, then masks it in one click — all without sending anything to a server.
What the AI does
- Runs an OCR model in the browser to read text in screenshots and photos
- Applies regex rules tuned for emails, phone numbers, postcodes, financial identifiers, government IDs, and credentials
- Runs a small BERT-based named-entity model to flag person names
- Uses spatial reasoning to extend masks across multi-line addresses
What the AI does not do
- Send anything to a third-party API
- Train on your data
- Keep a copy of your image
- Make perfect decisions — manual review is still recommended before sharing
How to use the AI
- Drop, paste, or open an image.
- Click Detect text.
- Wait a few seconds for the model to run (faster on machines with WebGPU).
- Review the masks. Add or remove any.
- Download the cleaned image.
Why on-device AI matters here
The whole point of redacting an image is to not expose it. Sending it to a remote AI service would defeat that. Redacted ships the OCR and PII models as part of the app, so the image stays on your device from open to download.
Privacy
OCR model: runs in WebAssembly (or WebGPU on supported browsers). PII detection: small ONNX model loaded into the browser. Neither sends data anywhere.
FAQ
What models are used?
A PaddleOCR detection + recognition pair for text reading, and a BERT-small PII NER model for names. All bundled into the app.
Does it use ChatGPT or another LLM?
No. The models are small and special-purpose, optimised for fast on-device inference.
How fast is it?
A typical screenshot processes in a couple of seconds on a modern laptop. Larger images and very dense text take longer. WebGPU acceleration speeds it up further when available.