Redactory

De-identification for AI training data early access

Sell the corpus, not the people in it.

One person, one synthetic identity across every Slack message, ticket and email. A written determination the buyer's counsel can read. The mapping stays with you.

CSV clean
Reads from

Hosted or self-hosted · zero data retention · the mapping stays in your environment

  1. 01

    Upload

    Drop files in any format. Slack, Jira and Notion exports included.

  2. 02

    Redact

    Every name, email, phone and ID is replaced the same way in every file.

  3. 03

    Review

    Fix a miss once and it applies to every file. The engine also flags anything the judge wants a human to decide.

Redactory
Q3 payroll corpus · 3 files
3 reviewing

99.96%real-PII recall, measured
12+document formats supported
0 bytesretained. Zero-retention hosting, or run it on your own infrastructure.

How we redact

Six layers, one pass.

Every document runs the same sequence. Each layer catches what the one before it cannot describe.

01

Pattern layer

Structured identifiers never reach a model. Social security and card numbers, IBANs, API keys, IP addresses, phones and emails are matched deterministically, first, every time.

412-88-9107SSN 4024 0071 3392 1187CARD GB29 NWBK 6016 1331 92IBAN sk_live_9fJ2kLp0QwErKEY 10.4.22.9IP +1 617 555 0148PHONE t.herrera@acme.comEMAIL 1987-03-14DOB
02

Language model

A NER model reads context and finds the names, companies and addresses no pattern can describe. It runs on our infrastructure, never a third party's.

Priya NairPERSON signed with Acme PressORG at 44 Fenwick RowADDRESS.

03

LLM as judge

Only the ambiguous hits go out, to a zero-retention endpoint. It votes keep, redact or review. It never rewrites your document.

“Dutch”
signature line, offer_letter.docx

keep redact review
04

Zero data retention

Nothing is stored once the run ends. Self-host it and the mapping stays in your environment. What leaves with you is the redacted set and the determination.

nothing kept
  • Upload
  • Redact
  • Determination
  • Source
05

Live collaboration

Several reviewers work the same corpus at once. A correction made once applies to every file, and the engine recalibrates on the next run.

3 reviewing
  • LA corrected “Dutch” across 2 filesnow
  • PN approved payroll-2026.xlsx12s
  • MD opened #finance-ops.slack.json40s
  • LA marked PROJ-142.jira.json as reviewed2m
  • Engine recalibrated, applied to 5 files2m
  • MD flagged a signature in offer_letter.docx6m
  • PN uploaded scanned-form.png9m
offer_letter.docxreviewed
payroll-2026.xlsxreviewed
#finance-ops.slack.jsonin review
PROJ-142.jira.jsonreviewed
scanned-form.pngqueued
06

Re-scanned before delivery

Every result goes back through the engine before you see it. Every corpus ships with a written determination: methods, recall on your sample, residual risk.

  • Second pass comes back clean
  • Every entity accounted for
  • Determination written and signed
Formats

Structured, unstructured, or scanned

Try it

Redact your documents

Drop a file and we will take you to pricing to pick a plan. Nothing you drop here is read, sent or stored. Hosted runs are deleted when they finish, or run Redactory on your own infrastructure.

New redaction run

no files
Drop files to redact instantly

or browse for files

Detecting and replacing sensitive data…

Before / after

Mode
Original
Loading…
Redacted
Loading…

Redacted files

What was replaced

OriginalReplacementType

Ready before the buyer asks.

Start with 500 documents. Nothing is retained, the mapping stays with you, and the determination goes in the data room.