AINPI

Open audit of the CMS National Provider Directory (NPD) public use files.

Live: https://ainpi.dev

Work in progress. AINPI is research and educational work. Data may be incomplete, stale, or wrong. Verify every number against primary sources before any business or clinical decision. The /insights page carries the full provenance analysis.

What it does

CMS publishes the National Provider Directory as FHIR R4 NDJSON files at directory.cms.gov: 21.7M records across 6 resource types (Practitioner, PractitionerRole, Organization, OrganizationAffiliation, Location, Endpoint) in the 2026-05-08 release. AINPI does three things with it.

1. Ingests the dataset into BigQuery and runs 31 pre-registered findings (H1 to H46). Some audit the directory itself. Others cross-check it against federal claims data: Medicaid spending, Medicare Part B and Part D, Open Payments, DMEPOS, and nursing-home ownership. 17 findings carry numbers today. 14 are registered and waiting.
2. Publishes every result as a static JSON file on a stable URL, with the null hypothesis and denominator registered before the numbers land.
3. Cross-audits federal exclusion lists (OIG LEIE and SAM.gov) against the directory. This closes 3 of the 4 federal database checks in 42 CFR 455.436. SSA-DMF stays restricted-access.

Each finding states what was measured, what it does not cover, and how to reproduce it. Where a result contradicts the registered prior, the page says so.

Recent findings

• H46, 2026-08-01. CMS lists a Medicaid provider directory for 32 of 51 jurisdictions. Five of those listed URLs do not resolve, so 27 of 51 have a catalogued directory the public can open.
• H45, 2026-07-13. Registered: the per-state gap between FHIR endpoints published through certified EHR vendors and what the NDH carries.
• H43, 2026-06-09. 99.98% of active practitioners carry a phone on the Practitioner record. The PractitionerRole to Location traversal returns nothing, which rejected the registered prior.
• H44, 2026-06-25. Of the 9 endpoint-metadata fields the HTE submission spec collects, 5 have no home in the NDH FHIR profile. Across all 114,071 FHIR-REST endpoints, no record uses the extensions that could carry the rest.
• H40, 2026-05-22. 194 federally-excluded NPIs billed Medicare Part B in the full window. Of 4 strict post-exclusion candidates, primary-source checks confirmed 1 and rejected 3 as SAM-NPI join errors.

Pages

Path	What it is
/	Landscape treemap. Every state and specialty cell scored on 6 audit dimensions.
/map	US choropleth of federally-excluded NPIs per state.
/findings	All 31 findings. Each registers its null hypothesis and denominator before numbers drop.
/npi	Per-NPI report cards for the high-risk cohort, with primary-source verify links.
/states	Per-state audit slices. All 50 states and DC.
/for-state-medicaid	Per-state pages written for Medicaid CMOs. Count and action first.
/methodology	Versioned methodology: DAMA DMBOK mapping, L0 to L7 scoring, reproduction commands.
/data-sources	Every dataset AINPI ingests, considers, or rejects, with license and refresh cadence.
/developer	API docs, code samples, and the MCP server.
/data-quality	D3 dashboard: choropleth, sankey, knowledge graph, drill-down, validation.
/insights	Provenance and variance analysis against published org numbers.
/provider-search	Live cross-source search across the NDH, NPPES, and 4 payer FHIR directories.
/npd	Public search by NPI, name, organization, state, city.

Public URL contract

Static JSON, CDN-cached, safe to depend on across releases:

• /api/v1/manifest.json discovery index: every finding and state URL, plus AI-agent tool schemas
• /api/v1/stats.json site-wide counters, methodology version, commit SHA
• /api/v1/findings/<slug>.json one per finding (types)
• /api/v1/states/<state>.json per-state slice, 50 states and DC

Breaking changes bump the path to /api/v2/. The shape never changes in place.

An MCP server at https://ainpi.dev/api/mcp exposes the same data as 5 tools.

Open data

Both ingested NDH releases are published as flattened parquet using the same extraction logic as the BigQuery tables, so numbers reproduce. directory.cms.gov serves only the latest release, so the archive is the only way to diff one release against another. Worked DuckDB examples that reproduce published findings live in examples/duckdb/.

What is in this repo

frontend/          Next.js 14 app: routes, API, charts, tests
analysis/          Python compute scripts, one per finding, output to /api/v1
crawler/           FHIR endpoint liveness crawler
docs/methodology/  Versioned methodology, rendered at /methodology
docs/reports/      Release updates, rendered at /reports/<slug>
docs/articles/     Long-form articles, rendered at /articles/<slug>
examples/duckdb/   SQL that reproduces published findings against the parquet
.github/           CI, CodeQL, secret scanning, weekly refresh, release

Architecture

       ┌────────────────────────────────┐
       │ directory.cms.gov              │
       │ manifest.json to 6 NDJSON.zst  │
       └──────────────┬─────────────────┘
                      │ analysis/fast_ingest_ndh.py
                      ▼
       ┌────────────────────────────────┐
       │ BigQuery (cms_npd dataset)     │
       │ resource:JSON + _* flat fields │
       │ 21.7M rows + 5 analytics views │
       └──────┬─────────────────────┬───┘
              │                     │
  analysis/   │                     │ scripts/sync-bq-to-supabase.ts
  h*.py       │                     │
              ▼                     ▼
       ┌──────────────┐     ┌──────────────────┐
       │ static JSON  │     │ Supabase Postgres│
       │ /api/v1/*    │     │ Prisma ORM       │
       │ on Vercel    │     │ pre-agg metrics  │
       └──────┬───────┘     └──────────────────┘
              │
              ▼
       ┌────────────────────────────────┐
       │ React + D3, static pages       │
       └────────────────────────────────┘

Why this split. BigQuery holds 40 GB of FHIR JSON for under $1 a month and every query runs under a hard byte cap. Findings compute offline and publish as static JSON, so a page view never triggers a paid query. Supabase holds auth and pre-aggregated metrics.

Quickstart

cd frontend
npm install
cp .env.example .env.local   # fill in Supabase + GCP values
npm run db:push              # push Prisma schema to Supabase
npm run dev                  # http://localhost:3000

To reload the warehouse, which is only needed when CMS publishes a release:

python analysis/fast_ingest_ndh.py --print-manifest-only   # check for a new release
python analysis/fast_ingest_ndh.py                         # download and load
npm run bq:sync                                            # aggregate to Supabase

Testing

npm run test         # Vitest, 175 unit tests
npm run test:e2e     # Playwright, 5 specs

Covers FHIR reference extraction, API parameter parsing, the validation contract, filter-context hierarchy, NPI and URL regex, BigQuery schema, the findings-hub data layer, and search.

Documentation

• CLAUDE.md architecture and developer reference
• DATABASE_SETUP.md Supabase, Prisma, and BigQuery setup

Sibling repositories

Repo	Scope
FHIR-IQ/ainpi-probe	FHIR endpoint liveness crawler (L0 to L7). Runs separately from the site so operators can audit the code that hits their endpoints.
FHIR-IQ/ainpi-examples	Python and DuckDB examples for the /api/v1/* contract.

Key references

• CMS National Provider Directory
• HTE Data Release Specifications
• NDH FHIR IG STU1 v1.0.0, the published spec. The STU2 CI build tracks upcoming changes and is not a stable reference.