Inspiration

"Has my order been paid?" is a short question with a lot riding on the answer. A small seller needs to give the buyer a clear reply without promising something the order details do not support. Seller Order Helper brings the question and the PayPal order details together, then helps the seller prepare an answer.

What it does

Start with a sample buyer payment question. The app creates a $10 order in PayPal Sandbox, reads that order back, and shows its details beside a reply draft in a local browser page.

The seller can review the wording and suggested next step before deciding what to tell the buyer. The app does not send messages or move money. The demo uses PayPal's test environment.

How I built it

Python connects to PayPal using OAuth and the Orders API. IBM Granite 4.2 3B runs locally through Ollama and chooses a response type and tone from a small set of options. The application then fills in the amount and order status from PayPal.

This gives each part a clear job: PayPal supplies the order details, Granite helps shape the response, and the seller makes the final decision. The review page puts those pieces in one place.

Challenges

The hardest part was making the answer useful without overstating what happened. A newly created order does not mean the buyer has paid. The wording has to reflect that clearly, even when the buyer asks a simple yes-or-no question.

I kept the payment details under application control and gave the model a focused role. That helped keep the reply tied to the order while still making room for different wording and tones.

Accomplishments

The connected flow has successfully created and read a PayPal Sandbox order, used the local Granite model, and produced a reply draft for the review screen. The source is public under the MIT license, with instructions for running it locally.

What I learned

A useful AI feature can have a small, specific job. For this project, the important work was deciding which facts the reply could safely use and keeping the seller in control of what happens next.

What's next

The next useful step is to try the workflow with sellers, learn which payment questions come up most often, and make the local setup easier. The current prototype keeps the scope focused on one order and one reply.

AI tools used

IBM Granite runs inside the application. Codex/ChatGPT helped with implementation and testing, Claude provided code reviews, and Qwen Coder helped with separate Windows code checks during development.

Built With

Share this project:

Updates

Submission history