NuvoVet

The clinical data layer between veterinary clinics and pet insurers (Korea first).

NuvoVet turns a messy veterinary invoice into a standard-coded, clinically checked insurance claim:

• Insurers get structured claims, a computed payout and explainable findings. Findings cover coverage, clinical plausibility, pricing and integrity.
• Clinics get a free pre-check before the owner files.

Why we pivoted from a clinic-paid drug-interaction checker, and the plan from here: docs/strategy/pivot_memo_2026-10.md.

All claims, clinics and patients in the demo are synthetic. Price benchmarks are seed estimates unless a public source is cited.

Products

Two products under one master brand, each with its own colour (frontend/src/brand/): nuvovet DUR (teal, clinic side) and nuvovet Claims (cobalt, insurer side).

Route	Product	What it does
/	both	Landing: a 3D laptop replays each product (EMR + DUR island, Claims adjudication), role chooser, one section per product
/dur#/emr/V1	DUR	Fictional EMR with the DUR overlay: the DUR island (draggable pill that peeks open on new findings, expands into the full review, docks into a side panel), row badges, the save gate, a guided 4-step demo strip; 10 fictional patients
/dur	DUR	Case studies, workbench, printable report and owner handout for the same engine
/insurance	Claims	Claims console on the live engine: review queue, live claim composer, clinic risk, fee benchmarks, engine performance, API docs
/clinic/claim	Claims	Free clinic-side pre-check with receipt-photo extraction

Patient photos: 나비 is a bundled CC0 photo (scikit-image's "Chelsea"); the other patients' photos are hot-linked from the Unsplash CDN (frontend/src/brand/pets.js), with a species-glyph fallback offline.

How the claims engine works

receipt photo ──► extract (Claude, transcription only) ──► Claim
                                                            │
       codebook: invoice text → NVP procedure / NVD diagnosis codes
       knowledge: Korean brand/ingredient → drug, therapeutic class, dose refs
                                                            │
     ┌──────────── rules (deterministic, every finding carries evidence) ───────────┐
     │ coverage   waiting period · exclusions · non-medical items · payable         │
     │ clinical   procedure↔diagnosis · undisclosed chronic condition · species     │
     │            safety · NSAID combos · dose vs reference                         │
     │ pricing    clinic's posted fee · regional percentile · quantity              │
     │ integrity  duplicates · species/identity mismatch · time bar                 │
     └──────────────────────────────────────────────────────────────────────────────┘
                                                            │
                     decision: auto_approve | review | deny_recommended
                     (never auto-denies; an adjuster confirms)

Code lives in backend/claims/. The API is in backend/routers/claims.py.

Run it

# backend (no database needed; drugs load from local JSONL)
cd backend
pip install -r requirements.txt
uvicorn main:app --reload --port 8000
python -m pytest -q tests            # 51 tests incl. synthetic recall gate

# frontend
cd frontend
npm ci
VITE_API_URL=http://localhost:8000 npm run dev

Optional:

• ANTHROPIC_API_KEY enables receipt extraction (POST /api/claims/extract).
• DATABASE_URL enables accounts and patient records.
• NUVOVET_ENV=production requires NUVOVET_SECRET_KEY.
• NUVOVET_ADMIN_PASSWORD seeds an admin account.
• NUVOVET_CORS_ORIGINS restricts CORS.

The console falls back to a static snapshot of the synthetic batch when the API is unreachable. Regenerate it after engine changes:

cd backend && python scripts/export_claims_demo.py

API

Method	Path	Purpose
POST	/api/claims/adjudicate	claim (+ policy, history) → decision, payable, findings
POST	/api/claims/precheck	clinic-side check before submission
POST	/api/claims/extract	receipt image → claim draft
GET	/api/claims/demo, /api/claims/demo/{id}	adjudicated synthetic batch
GET	/api/claims/evaluation	recall / false-alarm on labelled synthetic claims
GET	/api/claims/codes	procedure and diagnosis codebooks
GET	/api/drugs/search	drug search (Korean and English)

Data provenance (read before any commercial use)

See docs/DATA_PROVENANCE.md. In short:

• The legacy drug records in backend/data/converted/ were LLM-converted, largely from Plumb's Veterinary Drugs, whose licence forbids derivative works and AI use.
• They must be replaced by a clean-room rebuild before commercial use.
• The claims engine's own knowledge files are NuvoVet-authored, with public references. These are the therapeutic classes, Korean aliases, clinical rules and codebooks.