End-to-end wizard: Dashboard (practices completed in this browser, localStorage) → NPI lookup (live NPPES + mock demo NPIs) → confirm practice → AI-guided configuration (6 areas: demographics, prescribing, prior auth, samples, coverage, provider roles) → summary → return to Dashboard. React + Vite frontend; FastAPI backend. LangGraph agent: classify → validate (reply quality + sidebar_caption) → respond → check_complete. Qwen 3 (Groq) classifies phase and validates replies when GROQ_API_KEY is set; Claude Sonnet generates replies. Langfuse hosts the system prompt (onboarding-system-prompt, production label). Sidebar shows settings with intelligent captions; completed rows are clickable to revise. Escalation (multi-site / EMR, or repeated nonsense / validator suggestion) shows a specialist banner; chat continues for areas the AI can still configure.
~2.3 CS calls and ~45 min per practice → ~60 CS-hours/week at current volume. Every practice hits configuration before PA, samples, or drug tools—this is where CS time concentrates. Proving AI can configure a standard practice from NPPES + conversation tests the core bet: break the linear link between signups and CS labor.
uv sync && uv run uvicorn app.main:app --reload --port 8080 --env-file .env
cd frontend && npm i && npm run dev # second terminal → http://localhost:5173
# Optional: `frontend/.env` → `VITE_SPECIALIST_PHONE=+15551234567` for Schedule Walkthrough `tel:` on the summary (escalation uses built-in Call +17743579384)Demo NPIs: 1234567890 · 9876543210 · 5551234567 (mock if NPPES misses).
Env: Copy .env.example → .env. Required: ANTHROPIC_API_KEY. Optional: ANTHROPIC_MODEL, GROQ_API_KEY (phase classify + reply validate; else keyword + heuristics), LANGFUSE_PUBLIC_KEY / LANGFUSE_SECRET_KEY (prompts + observability).
Tests: uv run pytest tests/ -v — overview of modules and commands in tests/README.md. Optional live-LLM chat cases in tests/test_api_e2e.py skip without ANTHROPIC_API_KEY. Functional E2E (NPI + chat, completion vs escalation, mocked Claude): tests/test_functional_e2e.py; manual browser scripts in docs/e2e-scenarios.md. Agent/dev guide: AGENTS.md.
- Shadow 20 onboardings—AI vs CS config match rate.
- Guided first real workflow (e.g. one PA) post-setup—activation, not just setup.
- Instrument first real workflow within 48h, not completion rate alone.
FastAPI · uv · LangGraph · Claude (Anthropic) · Qwen/Groq · Langfuse · NPPES · React/TS/Vite · pytest · Docker
Product brief (aligned with shipped behavior): docs/product-brief.md
Diagrams: docs/architecture.mermaid, docs/user-flow.mermaid.