Quirkbook

The information guests need already exists: it's in the manuals and in the host's head. What was missing was a way to ask. So the manuals become something you can talk to.

• Photograph a rating plate and the manual gets found for you
• The house interviews its own host, one question at a time, instead of handing over a form
• It drafts its own voice, greeting and local guide, and the host approves them
• Guests scan a QR code and chat in their own language, no app to install
• The same knowledge becomes a host-reviewed manual, on the web and as a print PDF

• Vision reads the model number off a photo, and search finds the matching manual, so setup starts with a camera and not a form
• An agent is generated per property, from its manuals and its host's answers
• A hard rule that the house never invents appliance instructions
• One knowledge base feeds both the agent and the print manual through headless Chromium, so the two can't disagree
• Row-level security keeps each property's knowledge to itself

• The interview runs backwards: the product asks the owner the questions
• Every house ends up with its own personality
• The pilot house's agent knows the kayak, the water and how to light the BBQ

About eleven weeks, on and off, around a real house.

11 weeks · 104 commits · 1 pilot house

Trust was the whole problem. A wrong answer about a gas appliance isn't a funny chatbot mistake, so the house never invents appliance instructions. The interview was the other one: it had to feel like a conversation, one question at a time, not a form.

1. Models

   • Gemini Flash Lite

2. Build

   • Next.js 15
   • TypeScript
   • Tailwind CSS v4
   • Headless ChromiumPDF

3. Data and hosting

   • Vercel AI Gateway
   • SupabasePostgres, Auth, RLS

4. Watch and earn

   • Stripe