Noe: Autonomous Commerce Operations

• 

  Noe answers WhatsApp customers and updates operations from one owner console.

Inspiration

I wanted to give a small business something closer to a reliable virtual employee than a generic chatbot. I built Noe to answer customers on WhatsApp, remember context, and turn conversations into useful work without losing the owner's control.

What it does

Noe handles a seeded WhatsApp-style conversation queue in autonomous or supervised mode. In the end-to-end scenario, Maya's request evolves from a product question into a jacket and tote order, a Saturday pickup, inventory reservation, and a clearly marked non-fiscal invoice PDF. Noe answers product and availability questions, keeps the order up to date, and routes guarded or uncertain turns to the owner for approval.

How I built it

The React UI talks to a Node server-side demo engine through /api/demo. A state machine and tool layer own catalog, orders, approvals, events, and document transitions. I use Gemini through Google's GenAI SDK in the isolated Cloud Run runtime; local fixture mode is explicit and deterministic for reproducible development. Structured model output is validated before mutations, and an independent safety policy can hold suspicious replies. State snapshots use atomic temp-file plus rename persistence.

Challenges and learnings

The hard part for me was making an agentic assistant trustworthy enough to act like an employee without hiding its decisions. I learned to make every side effect visible, keep failure retryable, and separate the live-model boundary from deterministic fixture mode. The result is a complete workflow a reviewer can inspect from customer message to operational outcome.

Scope

I intentionally kept this as a self-contained hackathon demo. The customer transport is simulated to model the WhatsApp experience; no real WhatsApp message, billing, fiscal issuance, or commercial data integration is performed.

Built With

• cloud
• gemini
• google
• node.js
• react
• typescript
• vite