SkinDeepRESEARCHSteve Seguin

Use case · Related prototype

Find text to redact

Mark spans instead of writing a new document.

  1. Original message
  2. Locate name spans
  3. Review highlighted text
Illustration of the proposed workflow.

A support tool could highlight names before a transcript is shared. Returning exact text spans lets the application show what would be removed before changing the document.

Where the small model fits

A model assigns labels to tokens or character spans. The application applies redaction using those locations rather than asking a language model to rewrite the entire message.

What would need to work

Measure missed names and incorrect removals. The current prototype covers names; it is not a complete detector for addresses, account numbers or every kind of personal information.

Related experiments and evidence