-
-
PayPal sandbox checkout, opened only after human approval.
-
Top alternatives and the reason each rejected option lost.
-
Describe your purchase in plain language. The agent starts fetching vendor quotes through MCP
-
Best bundle within budget: 497.88 dollars of 500 used, with a 4.65 average rating
-
Vendor price comparison. The selected items are highlighted in green
-
All candidate quotes in one table. You can group by vendor or export to CSV.
-
Order confirmed through the PayPal Orders API, with the capture ID and purchased items.
Inspiration
Small businesses lose hours comparing vendors, prices, shipping and ratings before every purchase. We wanted one agent that does the comparison and still keeps a human in control of the payment.
What it does
A user types what they need and their budget in plain language, for example: "We have a 500 dollar budget. Find 5 mechanical keyboards and 5 wireless mice from our approved vendors." ProcureAI then:
- Uses Google Gemini to turn the request into structured requirements
- Fetches quotes from three vendors through an MCP (Model Context Protocol) vendor server
- Picks the best bundle with a deterministic optimizer (within budget, highest rating that fits)
- Explains the choice in plain language, with a trade-off analysis, a vendor price comparison, top alternatives and the reason each option was rejected
- Asks for human approval, then pays through PayPal Checkout (sandbox)
- Shows the confirmation and saves an order history, with CSV export of all quotes
How we built it
Node.js and Express on the backend, React with Vite and AG Grid on the front end. The vendor catalog is mock data served by an MCP server. The PayPal Orders API v2 creates the order from the optimizer's total and captures it after approval. Gemini (gemini-3.5-flash, with gemini-flash-latest as fallback) handles request parsing and explanations. The app is deployed on Render.
Challenges
- Gemini sometimes returns 503 "high demand" errors, so we added a fallback model, a regex parser and template explanations so the demo never breaks.
- Preventing duplicate checkouts and handling the PayPal return and cancel flows.
- Serverless platforms could not run our MCP child process, so we deployed on Render instead.
Accomplishments
A full flow from a plain-language request to a captured PayPal sandbox payment, with a human approval step and clear explanations.
What we learned
How to combine LLM parsing with deterministic optimization, and how to integrate the PayPal Orders API safely.
What's next
Real vendor integrations, multi-currency support, approval workflows for teams, and a persistent database.
Built With
- ag-grid
- express.js
- github
- google-gemini-api
- javascript
- model-context-protocol
- node.js
- paypal-orders-api
- paypal-sandbox
- react
- render
- vite
- zod