Zero-PHI-Egress Guardrails for Clinical LLM Applications
clinical-guard sits in front of clinical reasoning LLMs to inspect every incoming prompt locally on CPU before any external API call is made. It combines fast semantic crisis detection, hybrid prompt-injection defense, and ONNX PII anonymization with deterministic role-based routing.
Safety classification, injection screening, and PII redaction run 100% locally in-cluster. Sensitive patient identifiers never leave your VPC during pre-filtering.
Distinguishes layman from medical_professional. Overdose/toxicology queries trigger emergency hotlines for patients, but route to clinical reasoning (with audit flags) for clinicians.
Cosine-margin thresholds separate acute crises (hard_block) from ambiguous medication questions (soft_flag), preventing unnecessary false blocks.
4-Container Microservice Architecture
FastAPI gateway, concurrent pipeline orchestration, deterministic policy matrix, and Showcase UI.
multilingual-e5-small CPU worker for EN/DE crisis & injection anchor similarity.
Llama-Prompt-Guard-2-86M + delimiter smuggling regex worker.
multilang-pii-ner-ONNX + regex hybrid PII detector & entity table.
Model Selection & Benchmark Explorer
Click any benchmark card to inspect per-row predictions, misclassifications, and EN/DE breakdowns from GET /v1/benchmarks.