Inspiration

Shopping for several items within a budget can mean translating preferences into filters, comparing products and checking the final cost. I built ClearCart for shoppers who want to describe a task in ordinary language and keep control over what they buy. The prototype explores that experience with a small, inspectable catalog and visible totals.

What it does

ClearCart turns a request such as “A lightweight travel kit under $75, no leather” into a checked basket. Google Gemini interprets custom shopping requests and recommends preferred catalog products. The three built-in starter tasks have fixed category lists, so optional suggestions cannot become required items. Server code matches the fictional catalog and independently enforces the budget, materials, quantities and stock. The shopper sees the items, shipping, demonstration tax and complete total before explicitly approving checkout.

The demonstrated basket contains an Everyday Daypack ($27), Steel Water Bottle ($18) and Compact Travel Charger ($24). The $69 subtotal plus $5 shipping and zero demonstration tax totals $74. The buyer approves on PayPal's sandbox page. ClearCart captures the order, retrieves the payment and verifies the completed capture's identity, exact USD amount and currency before saving a receipt. All funds and catalog products in this MVP are fictional.

How I built it

I built the browser app with Next.js, React and TypeScript. Zod validates inputs and structured AI output. Node's built-in SQLite preserves server-issued quotes and payment records. Direct HTTP adapters connect Google Gemini and PayPal Orders v2.

AI supplies intent, while server-owned catalog prices determine the payable amount. A browser session owns each quote, which expires after 15 minutes. Persisted provider request IDs and the order ledger make retries reuse the same receipt. Controlled tests cover invalid requests, constraints, session ownership, mismatched payment evidence and duplicate callbacks.

Challenges and progress

The key challenge was connecting flexible language to a payment flow with precise, inspectable rules. I also had to resolve a locked sandbox merchant account and verify the replacement account's actual create, approval and capture flow. Genuine Gemini interpretation and a completed $74 USD PayPal sandbox capture now work locally. All 181 automated tests passed, including checks for starter category drift, custom extra items and provider errors. A fresh public clone passed type checking, the production build and 11 sample-mode HTTP judging checks. The 390-pixel mobile check showed no horizontal overflow. A 95-second captioned demonstration is public on YouTube.

What comes next

ClearCart currently uses 12 fictional products, USD, flat $5 shipping and zero demonstration tax. It supports one local Node process and does not provide real fulfillment. I have not measured shopping-time or conversion benefits. Next steps are short user tasks, accessibility feedback and merchant-supplied catalog/shipping data, followed by broader currency and deployment support.

Built by Moses Fawole, a solo entrant based in Nigeria, with Codex assistance.

Tools and AI disclosure

  • Google Gemini, gemini-3.5-flash-lite: actual structured-output calls interpret the shopping request into constrained intent. The exact built-in starter tasks use server-owned category lists; custom requests retain Gemini's interpretation. Independent server validation follows; Gemini cannot set prices or certify payment. The running app's public configuration confirmed this model on 3 November 2026.
  • PayPal Orders REST API v2, sandbox: creates orders for saved quotes; the buyer approves; server captures, retrieves and checks completed payment evidence. No live endpoint or real-money purchase is supported.
  • Codex: assisted with product planning, application implementation, controlled tests, review, verification and documentation. Moses remains the solo entrant.
  • Next.js, React, TypeScript, Zod, Node 24 built-in SQLite: browser UI, server routes, input/output validation and local persistence. Vitest: controlled automated verification. Lucide and CSS: catalog category icons and visual presentation. No AI-generated product photography or external merchant catalog is used.
Share this project:

Updates

Submission history