Show HN: VeilPhantom – Open-source on-device PII detection for AI pipelines
Hacker News Show AI
•
Generative AI
NLP
AI Research
I built VeilPhantom, a Python SDK that detects and tokenizes PII before text reaches any LLM. The problem: AI agents processing meetings, emails, tickets are handling raw sensitive data. Names, salaries, medical details - all flowing through cloud APIs. The solution: Detect PII on-device, replace with tokens ([PERSON_1], [AMOUNT_1]), send safe tokens to LLM, rehydrate response locally. Interesting finding: In benchmarks (98 scenarios, 8 verticals, Claude Haiku), accuracy went UP with PII redaction - 91.5% → 93.3%. Token-structured input seems to help models parse arguments reliably.