Presidio
Microsoft's open-source. NER + regex + validators. Handles emails, phone, SSN, credit cards, medical IDs. 40+ entity types.
Advertisement
LLM verification
Regex catches obvious. LLM catches contextual (names, addresses in prose). Two-stage: regex fast pass + LLM contextual.
Advertisement
Redaction strategies
Full: replace with [REDACTED]. Hash: SHA-256 + salt (allows joins). Format-preserving encryption: keeps schema. Pseudonymize: consistent aliases.
Restore for user
User expects their real name in response. Store mapping session-local. Substitute back in output. Never store cross-session.