Upload public records → Claude extracts entities → 17 fraud-detection rules fire → citation-bearing referral package for the AG / IRS / FBI.
Built backwards from a real Ohio nonprofit investigation. Five agency referrals filed.
cp .env.example .env # fill in DJANGO_SECRET_KEY and ANTHROPIC_API_KEY
docker compose up -d
docker compose exec backend python manage.py seed_demoOpen http://localhost:5174 — a complete investigation case loads automatically.
 |
 |
| Web — Cytoscape.js entity graph showing knots (persons + orgs) and connections. Click any node to drill into relationships, financials, and angles. | Research — Pull IRS 990 filings, Ohio SOS records, county recorder deeds, AOS findings, and statewide parcels directly into the case. |
 |
 |
| Financials — Multi-year 990 data in one view: revenue trends, officer compensation, balance sheet. Parsed from IRS TEOS XML via HTTP range requests — no third-party APIs. | Referrals — Deterministic, citation-bearing export. Every sentence traces to a source document. SHA-256 chain of custody on every file. |
Upload PDF → SHA-256 hash → OCR / extract text → Entities resolved
→ 17 signal rules fire → Investigator reviews → AI pattern analysis
→ Citation-bearing referral package
The deliverable is a deterministic referral package — not an AI summary. AI findings cap at evidence_weight=DIRECTIONAL; the investigator promotes them after verification. Generated text never reaches the file an agency reads without explicit human confirmation.
1. Audit-first data model. SHA-256 chain of custody on every document, append-only audit logging on every mutation, immutable timestamp guards on government referral filing dates. Legal defensibility is a primary requirement of this domain, not a nice-to-have.
2. Human-in-the-loop entity resolution. Fuzzy matching surfaces candidates rather than silent-merging. An investigator must confirm before two records become one. A silent merge in an evidence chain is worse than an extra click.
3. AI as a triage aid, not a deliverable. Claude handles messy document extraction and pattern analysis. The referral package is template-driven and citation-bearing. Nothing AI-generated ships without human review.
4. Failure-isolated connectors. Each external source is its own module with its own tests. A 404 from one ArcGIS endpoint does not take down the IRS pipeline.
5. Backwards from a real case. Every signal rule, every data model field, every UI affordance traces to an actual pain point from the founding investigation. Nothing here is speculative.
6. The AI is held to an evidence bar — and it's measured, not promised. Every AI "Lead" is graded by a dedicated eval harness. Deterministic guards confirm each citation resolves to a real case document and the text carries no accusatory language; an LLM-as-judge then scores two axes against golden fixtures — faithfulness (is the Lead actually supported by the evidence it was given?) and overreach (does it assert a verdict as established fact instead of a pattern to review?). Most fixtures are negative controls — same-name-but-unrelated people, a fully-documented clean filing — where the correct output is zero Leads. Per-fixture thresholds gate the run. The credibility firewall this product depends on isn't a line in a prompt; it's a property under test.
AI-first, on purpose. Claude Code writes most of the implementation inside a harness this repo defines:
.claude/skills/:new-connectorscaffolds a failure-isolated public-records connector with its tests,smoke-testruns the live API health check,session-wrapcloses out a working session with the docs updated..github/workflows/: Claude reviews every PR (claude-code-review.yml) alongside CodeRabbit. CI gates the merge.- Strict TDD throughout. The failing test lands before the implementation does. 1,102 tests hold the line.
What ships in a referral package, how evidence is weighed, and when an entity merge is confirmed stay human decisions.
| Layer | Technology |
|---|---|
| Backend | Django 5.2 · PostgreSQL 16 · Django-Q2 async jobs |
| Frontend | React 18 · TypeScript · Vite · Cytoscape.js · D3 (timeline) |
| AI | Anthropic Claude API (Haiku + Sonnet) |
| Connectors | IRS TEOS 990 XML · Ohio SOS · Ohio AOS · 88-county Recorder · ODNR Parcels |
| Infrastructure | Railway · Docker · GitHub Actions CI |
1,102 backend tests covering connectors, API endpoints, all 17 signal rules, async job pipeline, AI pattern augmentation, upload pipeline, entity resolution, classification, data quality validators, and the referral PDF exporter. CI enforces the full suite on every push via a Postgres service container — no Railway-roulette.
# Run the full suite (inside Docker):
docker compose exec backend python manage.py test investigationsAI eval harness (investigations/tests/evals/) — a separate faithfulness/overreach suite that grades AI Leads against golden fixtures (see engineering decision #6). It runs in two lanes by design: the deterministic guards (citation resolves to a real document, no accusatory language) run inside the normal suite above, while the LLM-as-judge scoring is tagged @tag("eval") and excluded from CI — model calls are non-deterministic, so gating every push on them would make CI flaky. Run the judged evals on demand with a real key:
# Live AI evals (needs ANTHROPIC_API_KEY; writes a scorecard artifact):
docker compose exec backend python manage.py test investigations.tests.evals.test_lead_quality --tag=evalThis platform was built from a real public-records investigation into a nonprofit organization, conducted using only publicly available filings — IRS Form 990s, Secretary of State records, county recorder filings, audit reports. The investigation produced formal referrals to five federal and state agencies. Identifying details have been intentionally removed from this public repository; verification documentation is available on request.
Tyler Collins · GitHub · LinkedIn · tjcollinsku@gmail.com