Live system · MediGovern RAG

Governed AI for healthcare prior authorization.

A data-governed RAG pipeline that ingests policy PDFs, benefit tables and FHIR records, runs every document through a six-rule data-quality gate before embedding, and answers coverage questions with grounded citations, confidence scoring and a deliberate ABSTAIN path. Built around CMS-0057-F. All data is 100% synthetic — no PHI.

6DQ rule families
3Confidence bands
72hCMS urgent PA SLA · 2027
100%Synthetic · no PHI

Open the governance dashboard inside this page.

The live MediGovern dashboard — what the system knows, what it has quarantined and why — loads in a window over this page. Backed by a FastAPI service with Anthropic Claude and a pgvector knowledge base.

Living architecture

From policy document to a cited, auditable answer.

A guided, auto-playing walkthrough of the real pipeline — ingestion, the six-rule DQ gate, embedding into pgvector, grounded retrieval, confidence scoring, the ABSTAIN path and the immutable audit log.

medigovern · living-architecture ⤢ Fullscreen

Guided walkthrough

Follow one prior-auth question, end to end.

A 10-step, self-guided walkthrough that follows one coverage question — Maria's lumbar MRI — from governed documents to a cited, auditable answer (or an honest ABSTAIN), in plain English. Click through it, or open it fullscreen.

medigovern · step-by-step ⤢ Fullscreen

Why it matters

Governed AI, built for a regulator's question.

Proves

I build governed RAG, not "chat with a PDF". The governance layer lives underneath generation: a DQ gate before embedding, confidence computed from evidence, an explicit ABSTAIN on conflict, and full lineage on every query, chunk and answer.

Why it matters to enterprises

CMS-0057-F requires health plans to answer prior-authorization requests within 7 business days (2026) and urgent ones within 72 hours (2027), exposed over FHIR. This operationalises that obligation — and when a regulator asks "how did you decide?", the answer is already logged.

Governance controls

FHIR-aware ingestion Metadata extraction (CPT / ICD / payer / plan) Six-rule DQ gate (pass / warn / quarantine) Quarantine queue Grounded retrieval (pgvector) Computed confidence band ABSTAIN on conflict Immutable audit packet (JSON + HTML)

Pipeline

01

Ingest

Typed loaders for PDFs, CSV benefits and FHIR bundles — traceable to source.

02

Metadata

Extract type, IDs, dates, CPT/ICD, payer and plan.

03

DQ gate

Six rule families: pass, warn or quarantine — before embedding.

04

Embed

Chunk + bge-small embeddings into pgvector with a metadata index.

05

Retrieve

Metadata pre-filter, then semantic search grounded in policy text.

06

Answer

Cited answer + computed confidence, or ABSTAIN on conflict.

07

Audit

Immutable lineage record + audit packet for every query.

Built with

A free, reproducible stack.

Python 3.11 FastAPI SQLAlchemy Postgres + pgvector Anthropic Claude (LLM) fastembed / ONNX (no API key) reportlab · pypdf Docker Compose