Inspiration
Going through different market places made me realize that there is no guarantee of how well the product performs, so decided to created beta version of providing a fraud shield whenever a merchant is onboarded and its products are taken into consideration.
What it does
Our AI shopping agent evaluates products and merchants before you buy. It scores each product across different scenarios and trust signals, rates the merchant, and gives you a clear reliability score so you can decide with confidence. Once a product is selected, it automatically creates a PayPal order and hands you the approval link, so you go from “is this safe?” to checkout in one flow.
How we built it
We built the product as a multi-agent system using Hugging Face’s smolagents. Each agent has one narrow job:
Scoring and selection agent: evaluates products across different scenarios and trust signals, rates the merchant, and produces the product score and the reliability score shown to the user. Payment agent: takes the verified product (ID, title, price) and creates the PayPal order. It is a CodeAgent connected to the PayPal MCP server, and it calls create_order through MCP.
Challenges we ran into
A major challenge was to get the required output from my agent and in as minimum tokens as possible so that the llm calls can be reduced and the speed of response become faster.
Accomplishments that we're proud of
Getting proper scoring system on some manipulated data set, proper merchant scoring handling new merchants as well, reducing the token amount from millions per call to thousands.
What we learned
Learned the trade-off between using AI and pre-determined codes, architecting an entire app and deciding which features are important and which to focus on in a short time and also learned to read through tracing and improve agent functionality.
What's next for Ghost Buyer
Improving and complexifying the scoring system, fine tuning a pre-trained model specifically for analyzing this market. Providing a scoring system for customer to protect merchant from bad customers. Creating a merchant verifying system through DNS TXT and email verification.
Built With
- embedding
- fastapi
- huggingface
- llm
- next.js
- postgresql
- python
- restapi
- smolagent
- tools
- typescript
