Small model. No inference server.
Personal information.
On your terms.
A neural PII detector that runs in your browser. Inspect names, addresses, and other sensitive text without sending it to a server.
Try it with real documentsWebGPU Local inference
8 categories Context-aware detection
Experimental Not a safety guarantee
01 / Explore
Real text is messy.
These public-source excerpts were selected for variety, not model scores. Try them unchanged, edit them, or bring your own text.
Choose a document
One click to load. Nothing to upload.
Model prediction
Loading WebGPU
Loading the model…
Detected entities and character offsets
A prediction, not permission to share. The model can miss sensitive information and flag ordinary text. Unmarked text is not necessarily safe.
02 / Measure
Evidence before confidence.
Quality and speed answer different questions. See the misses as well as the matches, then measure inference on your own device.
03 / Under the hood
A small model.
A clear boundary.
Trained from scratch with PyTorch. The browser downloads the model, then WebGPU runs inference locally. A worker keeps the interface responsive.
Input text and predictions stay on this device. There is no inference API, analytics script, or text storage.
What it looks for
Names, email addresses, phone numbers, street addresses, birth dates, US Social Security numbers, payment cards, and IP addresses.
The initial scope is English text and common US formats. It does not cover every identifier, language, or document type.
Where it falls short
Entity boundaries can be wrong. Names can resemble ordinary words. Unfamiliar formats and missing context can cause missed detections.
Quality measurements use synthetic data, not an independent audit of private documents. Do not use this experiment as an automatic anonymization system.