De-Identifier

Take the student out of the record — before the record goes to an AI. Built for IEPs, evals, and everything the review team reads.

A KVEC staff tool

Two AIs run right here on this machine — records are never uploaded anywhere. After the first run, it all works with the Wi-Fi off. How it protects student privacy

  1. 1
    Add the recordsAn IEP, an eval report, a stack of either — Word, PDF, or scans.
  2. 2
    Two AIs read each oneDirect identifiers get scrubbed; contextual ones get underlined.
  3. 3
    You make the callsJudge the underlines against the neighbor test, then save.

Word .docx comes back as Word, formatting intact — including comments and tracked-changes authors, which get scrubbed from the file itself. .pdf and scans come back as text. Scans work when the record is typed — handwriting can't be read.

First use downloads both AIs once (about 650 MB — the contextual one is most of it) and keeps them on this laptop. A long document takes a few minutes; the page stays usable while it works.

How it works

  1. Feed it records. Paste one, or drop Word and PDF files anywhere on the page — a whole review stack at once is fine, each record is handled separately.
  2. Two AIs read it. The first scrubs direct identifiers — names, addresses, phones, birthdays, schools, ID numbers, plus the Word file's own hidden author fields. The second reads for context: diagnoses, medications, family members, churches, employers, teams, benefits. First use downloads both, once; after that it works with the Wi-Fi off.
  3. Judge the underlines. Contextual details come back underlined, not scrubbed — an AI usually needs the diagnosis to be useful, but in a small school “the seventh grader with a seizure disorder” can identify a child by itself. Click an underline to scrub it; a category chip scrubs the group.
  4. Send it to the AI. Copy the de-identified text into ChatGPT, Claude, or Gemini. Word files can also come back as real Word files, formatting intact.
  5. Put the names back. Paste the AI's reply into the box at the bottom of the review screen and the tags flip back to real names — on your device, like everything else.
The neighbor test — the standard this tool serves
  • De-identified means: a neighbor who knows the family couldn't recognize the student. Software handles names and numbers reliably; combinations are the reviewer's call — sport + grade + diagnosis can be a fingerprint in a 400-student school.
  • Before sharing, scan for: the only student who… (sport, instrument, disability, family situation), sibling references, church or team names, a parent's workplace.
  • No software can promise the neighbor test on its own. This tool's promise is narrower and honest: everything a machine can find is found, and every judgement call is made visible instead of silent.
Take it on the road (IEP reviews in no-signal buildings)
  • Install it like an app, once, on Wi-Fi: Chrome or Edge → install icon at the right end of the address bar (or this page's Install as an app button). Mac Safari: File → Add to Dock.
  • Run one de-identify while still on Wi-Fi so both AIs finish downloading, and wait for the green “✅ Road-ready” line on the front page. After that, airplane mode changes nothing.
  • Updates arrive by themselves next time the laptop is online.
Good to know
  • Nothing leaves the machine — records, names, findings, nothing. The privacy page says it in plain English, suitable for forwarding.
  • Word downloads are scrubbed inside and out: document text, headers/footers, comment authors, tracked-change authors, and the file's own Author property.
  • Records save under their tag (NAME-4-scrubbed.docx) with a WHO-IS-WHO key in batch zips — keep the key, never send it.
  • Grading essays instead? Use Paper Scrubber — same engine, lighter, built for teachers.