DeID Studio

Review the details.

Compare text. Correct detections. Export your decisions.

Connecting to analyser
Synthetic text only. Never submit private records. Human review helps inspect model errors; it does not guarantee anonymity.
⌘ / Ctrl + Enter to analyse

Loading the server's synthetic examples…

01Source document

0 characters

Missed something? Select text in the Text view, then redact it locally.

02Output preview

Review required
No analysis yet.
See the transformation.

Analyse the synthetic source to preview suggestions. Then apply or keep each detection in Review.

Export ▴

No automatic storage. Exports happen only when you choose them.

03Detection review

Waiting for analysis
Make the call.

Confirm a transformation, keep a false positive, or add a missed span. Proposed detections start unreviewed.

Scope, policies and privacy

This prototype combines English statistical NER with prioritized identifier patterns. Both can miss or mislabel information. Proposed transformations come from the server's configured policy; apply/keep choices and manual redactions are assembled locally without another analysis request. Confirming detections does not verify the whole document.

Input and review choices stay in memory: no browser storage or analytics. Editing the source or changing policy clears the previous review. The server processes text in memory without deliberate storage; cancelled inference may still finish. Hashes use an ephemeral server salt and remain pseudonyms, not anonymous data. An export may contain identifiers you chose to keep or the detectors missed.

Output text exports include the current preview. Review JSON additionally contains labels, offsets and review status, but not a separate copy of the original source or matched values. Metadata-only copy excludes output text too.

English · CPU inference · Synthetic examples