HIPAA-Compliant Clinical Prior Authorization Evidence Agent (Capstone Project)
Reproducible healthcare AI capstone using Synthea synthetic FHIR records and public payer policies: compiles boolean medical-necessity criteria graphs, retrieves mixed FHIR/PDF evidence under PHI scrubbing rails, and gates releases via a 50-case clinical eval suite.
Best For
Candidates targeting Healthcare AI, Applied AI, or AI Engineer roles needing to prove mastery of regulated, audit-grade agent systems over synthetic FHIR data.
Human-AI Boundary
Renders Met / Evidence Missing / Clinician Attestation badges with one-click export of an annotated audit packet for physician signature.
Anti-Patterns to Avoid
- Training or testing capstone demos on real unmasked patient records from public scrapes
- Building a plain chat wrapper without deterministic checklist verification of ICD-10 / CPT and date/threshold rules
Latency/Cost Profile & Eval Metrics
Async case evaluation ~6-12s E2E on local/cloud OpenAI-compatible runtime; interactive HITL inspection < 150ms · Policy graph pre-compilation (once per policy) + per-case FHIR/PDF checklist verification
- Medical Necessity Criterion Sensitivity & Specificity: >= 0.94 sensitivity / zero false-Met endorsements on missing prerequisite labs (Evaluated across 50 synthetic cases (25 eligible, 15 incomplete, 10 edge conflicts))
- PHI Leak Redaction Precision: 100% synthetic MRN/DOB/SSN redaction before trace export (Verified by automated Presidio/regex telemetry interceptor tests)
💰 Business ROI Model & Evidence Level
portfolio_blueprint🏗️ End-to-End Architecture Pipeline
Load structured FHIR resources and parse synthetic fax PDFs into coordinate-indexed chunks.
Run LangGraph per-criterion verification with mandatory Presidio PHI scrubbing prior to trace logging.
Benchmark accuracy delta vs naive prompting and expose clickable page anchors for clinician attestation.
📦 Project Deliverables & Portfolio Artifacts
- Policy-to-Criteria Graph Compiler — Compiles public payer coverage PDFs into versioned AND/OR checklist JSON trees
- Zero-PHI Telemetry & 50-Case Eval Harness — Includes PHI redaction unit tests, missing-evidence adversarial cases, and Ragas regression CI
🎤 Interview Defense & System Design Deep-Dive Q&A
- When comparing a single 100k-token prompt containing full patient history + policy vs your Criteria DAG architecture, what were the exact false-positive failure modes?
Bidirectional Taxonomy Links
Internal Tools & Primary Sources
- Synthea Synthetic Patient Generator (MITRE Open Source) (github, 2026-11-15)
- HL7 FHIR R4 Resource Specifications (official_docs, 2026-11-15)