The typed record the state machine builds as the chat progresses — this is what the dashboard and the portals read.

What's real here

• ✓ The conversation logic is the production build. intake.ts and the welcome copy are compiled straight from the repo — same prompts, same menus, same validation.
• ✓ Price guidance uses the demo valuation adapter — the same deterministic stand-in the server runs until LOOM is connected.
• × No database and no WhatsApp. The conversation lives in memory and resets on reload; the server build persists to Postgres and sends through the BSP.
• × The AI layer runs server-side, so it is not in this page. Two parts of the product call Claude from the API: the description writer, which drafts your listing copy for you to approve, and the concierge, which answers off-script questions — "how long does transfer take?", "who pays the rates clearance?" — in the same thread, and hands over to a registered practitioner when it should. Both need an API key, which has no business in a public web page. Run the API to see them live.

The scripted flow above is deliberately independent of all that: if the model is unavailable, every question here still works.