​
How it works

Every OCR line passes two layers:
  1. Rules with checksums: emails, phones, IBAN (mod-97), cards (Luhn), tax and national IDs, secrets after “password:” or “token:”, and person names (honorifics, “Name:” labels, Russian name forms).
  2. A PII model, run locally: GLiNER2 privacy filter. No outside service sees the unredacted text.
Matching boxes are blurred on the keyframe before it is stored or sent to a model.

​
Agent records

The same two layers run on the text values of every agent record a check sends to a model (judge, router) and on the record stored in the check log. Each value is checked with its field name (contact: Jo Doe), and matches become placeholders ([PERSON], [EMAIL]). Short values and identifiers get the rules layer; free text also goes to the local model when it is loaded. Precise rules still run on the original values, on the server. See Data sent to models.

​
What we measured

Caught: names, emails, phones, IBANs with and without spaces, card, SSN, German tax ID, Russian passport, INN, SNILS, date of birth.
Known gaps.
  • Missed a Russian street address.
  • Box edges can leave the first or last character visible.
  • The test screens were synthetic.
  • The evals ran with both layers. The hosted preview runs the rules layer only (to fit a small server), so expect lower recall there, especially for names and addresses.