[Feature] NER-based PII detection as alternative to regex patterns

[mosw]

Current PII detection uses pattern matching (regex) which catches structured formats like SSNs, credit card numbers, emails, and phone numbers. But it misses:

Names in context ("Send this to John Smith at Acme Corp")
Unstructured addresses
Context-dependent PII

Explore adding a lightweight NER (Named Entity Recognition) model as a post-processing step for PII detection in agent responses.
Considerations:

Keep dependency footprint light — evaluate spaCy vs small transformer models
Should be optional (users who want regex-only shouldn't need ML dependencies)
Could be a plugin/add-on: pip install agent-warden[ner]

Open questions:

What's the acceptable latency overhead per agent response?
How to handle false positive/negative tradeoff? Configurable sensitivity levels?

[NeuZhou]

We implemented regex-based PII detection in ClawGuard (https://github.com/NeuZhou/clawguard) covering emails, SSNs, credit cards, API keys, phone numbers, etc. All 100% local, zero cloud. Would be interested in comparing approaches — NER would catch more edge cases but adds model dependency. Perhaps a hybrid approach?