Clinics want cloud AI to write discharge advice in Hindi. The note carries Aadhaar, ABHA and phone numbers, names and home addresses. This agent strips them inside your browser, sends the cloud model only placeholders, checks its answer, and puts the real values back on your side.
Run the agent to see the answer with the real values back in place.
Nothing sent yet.
| Loading measurements… |
Protection recall counts an identifier as protected only when every character of it is covered. Precision counts a redaction as correct when it overlaps a labelled identifier.
All test notes are synthetic. Each test set was written by a separate author who never saw the detection code. Blind set 2 is the honest measure for the current rules; set 1 was used to improve them after its first, blind run. Real notes will be messier, so treat these numbers as an upper bound until a clinic pilot measures leaks on real, consented text.
Redaction runs in your browser through Pyodide. Only the redacted note crosses the perimeter. The demo endpoint re-checks for identifiers and refuses text that still contains them, caches answers, and stops after a small daily number of model calls.