Clinical Co-Pilot + Adversarial Eval Platform
What it does
A FastAPI clinical agent with an OpenEMR module, paired with a continuous adversarial-evaluation platform that red-teams the deployed agent on its own.
- Each attempt is judged by deterministic-first oracles (canary match, tool-call audit, log scan, token counter, HTTP assertion). An LLM judge is used only where a result can’t be checked in code.
- Confirmed exploits are frozen into a versioned regression suite.
- Every attack ships with a benign twin, so a “fix” that just breaks the workflow fails instead of passing.
- Attack + benign twineach attack has a control
- Clinical agentFastAPI in OpenEMR
- Deterministic oraclescanary, tool audit, logs
- LLM judgeonly where truly semantic
- Regression suiteconfirmed exploits frozen
Tools
Python, FastAPI, PHP (OpenEMR), Pydantic, pytest