Reads the documents behind a security-deposit insurance claim (lease, tenant ledger, move-out itemization, repair invoices) and recommends a payout, with every dollar traced to a document line. On a held-out test split it lands within $250 of the human adjudicator on 91% of claims, with a median error of $0, for about $0.33 a claim.
Screenshots show synthetic sample claims, no real claimant data.
- Claude Opus 4.8 reads the claim documents through a forced tool call and returns structured facts: each charge, the ledger balance, eligibility. A deterministic Python engine then applies the arithmetic, the policy cap, and the exclusions, so the payout is never a number the model made up.
- The engine anchors on the ledger balance instead of summing repair line items. That change cut mean absolute error from $909 to $121; guardrails and hybrid reading took it to $62.
- Each PDF page routes by text density: about 75% read free and lossless with pure-Python pdfminer.six, true scans go to Claude vision. No poppler or tesseract binaries, so the same code runs inside the packaged macOS desktop app (PyInstaller).
- Extractions are stored and the engine is a pure function, so a candidate rulebook (JSON,
config/rulebook.json) re-scores against the human decisions by replaying stored reads with zero API calls (python -m scripts.rescore). - The source CSV carries the human adjudicators' answers. An allow-list projection in
core/csv_data.pykeeps those columns out of the model input, and the boundary holds across the database round-trip; only the scoring code reads them. - Django-Ninja API over a framework-free
core/, React + Vite frontend served same-origin, one SQLite workspace per user. Deployed on Railway. No LLM framework, just direct Anthropic SDK calls validated at the boundary with Pydantic.
The full design is in ARCHITECTURE.md.
Requires Python 3.13 and an ANTHROPIC_API_KEY.
python -m venv .venv && source .venv/bin/activate
pip install -e .
echo "ANTHROPIC_API_KEY=sk-ant-..." > .env
python manage.py runserver # API + SPA at http://localhost:8000VITE_USE_MOCKS=1 runs the frontend on bundled synthetic fixtures, no backend and no key needed.