Inspiration

Small shops often receive purchase requests as messages: “Two notebooks and a desk mat, under $45.” Turning that sentence into a checkout normally requires another round of product selection, quantity entry, and price checking. IntentCart explores a short, understandable path from a conversational request to a payment the shopper explicitly approves.

What it does

IntentCart converts a natural-language request into a structured cart drawn from a small, fixed demo catalog. It displays the exact quantities, catalog prices, USD total, and optional budget check. The shopper reviews the cart, opens PayPal sandbox, approves a simulated payment with a sandbox buyer, and returns for server-side capture verification.

The AI prepares the order; the shopper decides whether to pay. Unknown products and requests exceeding the budget lead to clarification. Success is conditional on a completed PayPal capture whose amount, currency, and order reference match the reviewed draft.

How we built it

The prototype uses Python's standard library, a plain HTML/CSS/JavaScript interface, SQLite, a cloud DeepSeek model with JSON output, and PayPal's sandbox Orders v2 REST API. It runs locally with one command and needs no third-party Python packages, frontend build step, or local model.

The model returns product IDs and quantities. Python validates those fields and calculates prices with decimal arithmetic from a server-owned catalog. A stored draft ID connects the review screen to order creation. The PayPal integration obtains an OAuth token server-side, creates an order with a sandbox approval link, checks approval, captures the order, and verifies the returned receipt.

Challenges

The central challenge was making an AI-generated order trustworthy at the payment boundary. Model output is untrusted input, so the server rejects unknown products, price overrides, malformed quantities, and incomplete responses. Payment retries also need care: stable request IDs, session ownership, and persisted receipts prevent repeated clicks from creating a fresh payment request for the same draft.

Accomplishments

The complete flow has been exercised with a real DeepSeek request and actual PayPal sandbox buyer approvals and captures. The recorded English demonstration completed a $42.00 USD payment on November 9, 2026, with order ID 2VF8317785897294A and capture ID 95802116K4171093R. The server verified the capture before rendering its receipt. Separately, all 18 automated contract tests passed; those tests use mocks and are not represented as payment evidence.

What we learned

Conversational checkout needs a clear handoff between understanding and authorization. A helpful assistant can translate intent while a deterministic order service owns prices and a human approves the final payment. A successful redirect alone is insufficient evidence of payment; the capture itself must be checked.

Scope

This submission is a deliberately focused local sandbox prototype with four fictional products, USD pricing, and one create–approve–capture flow. It has no physical fulfillment, live payment mode, shipping, taxes, or real-money checkout. The video uses actual browser captures with edited pauses and synthetic English narration. The fictional buyer and funding method were generated by PayPal for testing.

What's next

The next step beyond this hackathon prototype would be connecting the same review-and-approval pattern to a merchant-owned catalog and evaluating it with real shoppers. This is future work; the submitted scope remains the sandbox flow described above.

Share this project:

Updates

Submission history