Human approval before any delivery change
Pedaleando Ops Copilot is a last-mile delivery operations prototype for Argentina. It helps a merchant respond to a delay without letting an AI suggestion silently change a live order.
A delay creates an Incident. Ops Copilot then produces an AiProposal with a revised ETA and customer message. The operator can edit, approve, or reject it. Only approval updates the order; each decision is recorded as an audit event and synchronizes Merchant, Customer Tracking, and Courier views.
Why this matters
Delivery teams need fast incident handling, but customers and operators need a reliable human decision boundary. Pedaleando makes the workflow explicit: detect -> propose -> review -> approve -> synchronize.
Working demo
The public demo includes a live MapLibre route, multilingual ES-AR / EN / RU interfaces, light and dark themes, an operator-scoped Copilot panel, and a no-key deterministic fallback. The fallback intentionally produces only pending proposals, so the demo works without secrets or API billing.
The public MCP endpoint exposes exactly two read-only tools: get_delivery_context and build_review_link. It has no tool that can approve, apply, or mutate an order.
Evidence-aware AI cost view
The Merchant view checks official provider pricing sources and the official BCRA USD-ARS quote. The calculator labels values as Estimated API-equivalent cost. Quotes must be online_verified before they can influence an automated provider recommendation; changed, stale, or unavailable sources remain visible and fail closed.
Built with Codex and GPT-5.6
We used Codex with GPT-5.6 to design, implement, test, and harden the prototype. The product itself does not claim a runtime GPT API call: its demonstrated path is a deterministic, no-key fallback plus a read-only MCP boundary. This keeps the human approval contract demonstrable and reproducible for judges.
Run locally
npm ci
npm test
npm run build
npm run dev
The repository includes setup, MCP, pricing-source, accessibility, and QA documentation. The production endpoint provides /healthz, /api/version, /api/pricing, and /mcp.
Built With
- css
- gpt-5.6
- html
- javascript
- maplibre-gl-js
- model-context-protocol
- node.js
- openai-codex
- vercel
Log in or sign up for Devpost to join the conversation.