Scaminator

E-Commerce Fraud Detector
SENG 472 LLM Term Project

---

Overview

Fraud Detector is a multi-agent, LLM-powered system designed to assess the trustworthiness of e-commerce products and sellers. By analyzing product descriptions, reviews, and seller information, the system provides an overall verdict, a trust score, and actionable insights for users to make safer online purchases.

---

Features

• Multi-Agent Analysis:
  Separate agents investigate product descriptions, reviews, and seller profiles.

• LLM-Powered Reasoning:
  Uses Google Gemini and custom logic to generate human-like, explainable verdicts.

• Comprehensive Output:
  Provides a summary, trust score (0-100), detailed reasoning, and user suggestions.

• Modular Design:
  Easily extendable for new data sources or analysis agents.

---

Project Structure

fraud-detector/
│
├── backend/
│   ├── agents/
│   │   └── final_judge.py
│   ├── investigators/
│   │   └── comment_Investigator.py
│   ├── scraper/
│   │   └── product_comments.py
│   └── utils/
│       └── gemini_utils.py
│
├── frontend/
│   └── ... (UI code, e.g., Streamlit app)
│
├── requirements.txt
└── README.md

---

Installation

1. Clone the repository:

   git clone https://github.com/yourusername/fraud-detector.git
   cd fraud-detector

2. Install dependencies:

   pip install -r requirements.txt

3. Set up API keys:

   • Place your Google Gemini API key and any other required credentials in a .env file or as environment variables as needed by gemini_utils.py.

---

Usage

Backend

Run the backend analysis (example for product comments):

cd backend
python scraper/product_comments.py

Run the final judge agent:

python -m agents.final_judge

Frontend

If using Streamlit or another UI, run:

streamlit run frontend/app.py

---

How It Works

1. Scraping:
   Product and seller data are scraped from e-commerce platforms.

2. Agent Analysis:

   • Product Description Investigator: Analyzes the product description for suspicious patterns.
   • Product Reviews Investigator: Evaluates user reviews for authenticity and red flags.
   • Seller Information Investigator: Checks seller reputation and behavior.

3. Final Judge:
   Aggregates all agent outputs, generates a trust score, and provides a detailed, explainable verdict.

---

Example Output

### Summary of Analysis

The product description contains inconsistencies and several negative reviews mention non-delivery. The seller has a low reputation score.

# Likely Scam

### Overall Trustworthy Score: 18/100

### Overall Score Review

The product and seller both exhibit multiple red flags, including fake reviews and suspicious seller activity.

### Reasonings

**1.** Multiple reviews report non-delivery.  
**2.** Seller has only been active for 1 month.  
**3.** Product description contains grammatical errors.  
...

### Additional Suggestions for Users

Avoid purchasing from this seller. Look for verified sellers with a longer history and positive reviews.

---

License

This project is for academic use only. See LICENSE for details.

---

Authors

• Ahmet Alp MALKOÇ
• Utku ERDOĞANARAS
• Kutay BECERİR
• Berk BELHAN
• Alperen KARADAĞ

---

Acknowledgements

• Google Gemini API
• Streamlit
• Pydantic
• Selenium
• BeautifulSoup