LedgerLens

LedgerLens is a local-first fraud audit agent for mixed financial dossiers. It ingests GDPdU exports, CSV, XLSX, PDF, DOCX, and scanned PDFs; reconciles the accounting population; and promotes only evidence-backed, source-cited fraud hypotheses.

Quick start

cp .env.example .env
docker compose up --build

Open http://localhost:5173. The practice dossier is mounted read-only and loaded automatically on first start.

Local development

Backend:

cd backend
uv venv
uv pip install -e '.[dev]'
uv run uvicorn app.main:app --reload --port 8000

Frontend:

cd frontend
npm install
npm run dev

Design rules

• Original files are immutable and addressed by SHA-256.
• Calculations use decimal arithmetic; the LLM never performs reconciliation.
• Every finding includes native anchors and calculation lineage.
• Weak signals never create a finding without corroboration.
• Runtime rules do not contain practice vendor IDs, account IDs, dates, or amounts.
• If no model endpoint is configured, all deterministic analysis and audit chat remain available.

API

FastAPI documentation is available at http://localhost:8000/docs.

Tests

cd backend
uv run pytest -q

The sample regression suite validates all four seeded schemes and explicitly clears the seven decoys described by the sealed practice ground truth.

The updated practice labels are reflected correctly: the unrelated EUR 86,500 accrual is retained as counter-evidence and does not reduce the EUR 192,000 cut-off exception.

See docs/JUDGING_PLAYBOOK.md for final-dossier handling, auditor questions, and the demo sequence.