Hey PH 👋 Maker here. This started from a real incident — I accidentally sent a customer's credit card number to GPT-4o during a benchmark test. Saw it in the logs the next morning. Already gone. Spent the weekend building a firewall so it couldn't happen again. Every prompt gets scanned by Microsoft Presidio before it reaches any model. If PII is detected above a confidence threshold, the request is blocked — HTTP 403 — before it touches the wire. The part I'm most proud of: the policy engine is a YAML file. No restarts, no code changes — edit the file, hit one endpoint, rules are live. Ship a policies/hipaa.yaml or policies/gdpr.yaml to your team, and your compliance posture is a git commit. Happy to answer anything about the Presidio integration, the FastAPI dependency injection pattern, or the multi-provider routing. Ask away.

LLM Governance Engine

Self-hosted PII firewall for LLMs — policies, audit trail

Self-hosted PII firewall for LLMs — policies, audit trail