DEMM Listing Inspector Browser Extension

What it does

DEMM Listing Inspector is a Chrome extension that protects Etsy buyers from AI-generated product scams by analyzing multiple risk factors in real-time:

Dual-Layer AI Image Detection:

SynthID Watermark Detection: Uses Google DeepMind's SynthID technology to specifically detect Gemini-generated images Universal AI Detection: Leverages custom Gemini prompting to identify AI-generated images from ANY source (DALL-E, Midjourney, Stable Diffusion, etc.) by analyzing visual artifacts and patterns

Review Sentiment Analysis: Leverages Gemini API to analyze review text and detect sentiment-rating mismatches (e.g., negative reviews with 5-star ratings that may indicate fake or manipulated reviews) Visual Verification with CLIP: Uses OpenAI's CLIP (Contrastive Language-Image Pre-training) model to compare product showcase images against customer review photos, detecting visual discrepancies between advertised and actual products Custom Risk Scoring Algorithm: Our proprietary weighted algorithm intelligently combines multiple fraud indicators:

AI detection confidence levels Sentiment-rating mismatch severity CLIP image similarity scores Review pattern anomalies Produces clear, actionable risk assessments (Low, Medium, High, Critical)

Intelligent Caching: Uses Backboard.io's persistent memory API to cache analysis results, providing instant feedback on previously-scanned listings (reducing analysis time from 5-10 seconds to under 1 second)

How we built it

Frontend: • Chrome Extension with content scripts for seamless Etsy integration • Real-time data extraction from Etsy listings (images, reviews, seller info, pricing) • Interactive risk dashboard displayed directly on product pages Backend: • Flask API server for processing and analysis • Google Gemini API integration for SynthID watermark detection • OpenAI's CLIP model for semantic image similarity comparison between seller and buyer photos • TextBlob sentiment analysis for review authenticity checking • Backboard.io API integration for persistent caching across sessions • Comprehensive risk calculation algorithm weighing multiple fraud indicators Tech Stack: Python, Flask, Chrome Extension APIs, Google Gemini, OpenAI CLIP, Backboard.io, TextBlob

Challenges we ran into

CLIP Image Comparison: Implementing CLIP for cross-image similarity comparison was challenging. We had to handle varying image sizes, formats, and qualities while computing meaningful similarity scores between seller product photos and buyer review images. Backboard.io API Learning Curve: As a new platform, Backboard.io had limited documentation. We had to experiment with authentication headers (discovering it uses X-API-Key instead of standard Bearer tokens) and API response formats to successfully integrate their persistent memory system. Chrome Extension Permissions: Balancing powerful data extraction capabilities while respecting user privacy and Chrome's security policies required careful permission scoping. Real-time Performance: Processing multiple images through both SynthID and CLIP, plus analyzing dozens of reviews in real-time was computationally expensive. Implementing Backboard.io caching was crucial for achieving acceptable response times on repeat visits. Cross-Origin Requests: Coordinating data flow between the Chrome extension, Etsy's website, and our Flask backend while handling CORS properly. Model Selection and Testing: Iterating through multiple Gemini models to find the optimal combination for both SynthID watermark detection and general AI image analysis.

Accomplishments that we're proud of

Dual-layer AI Detection System: Developed a comprehensive AI detection approach combining:

SynthID watermark detection specifically for Gemini-generated images Custom Gemini prompt engineering to detect AI-generated images from ANY source (DALL-E, Midjourney, Stable Diffusion, etc.)

Custom Risk Scoring Algorithm: Designed and implemented our own weighted risk calculation algorithm that intelligently combines multiple fraud indicators:

AI detection confidence scores Sentiment-rating mismatches CLIP image similarity scores Review pattern anomalies Produces actionable risk levels (Low, Medium, High, Critical)

Multi-modal AI Analysis: Successfully integrated three different AI models (SynthID, CLIP, Gemini) into a unified fraud detection pipeline Semantic Image Comparison: Leveraged CLIP's vision-language understanding to automatically detect visual discrepancies between seller and buyer photos without manual review Production-ready Caching: Implemented Backboard.io's persistent memory API, demonstrating world-record 90.1% accuracy on long-term contextual memory benchmarks Real-world Impact: Created a tool that can genuinely protect consumers from financial loss and disappointment Seamless UX: Built an extension that works invisibly in the background and presents results naturally within Etsy's interface Etsy has become a hotbed for AI-generated product scams, where sellers use AI-generated images to misrepresent products, leading buyers to receive items that look nothing like what was advertised. We wanted to create a tool that empowers buyers to make informed purchasing decisions and avoid falling victim to these increasingly sophisticated scams.

What we learned

• AI Watermarking Technology: Gained deep understanding of how SynthID embeds imperceptible watermarks in AI-generated images and how to detect them programmatically • Vision-Language Models: Learned how CLIP's contrastive learning approach enables semantic image comparison across different photo styles and angles • LLM API Integration: Learned to work with Google's Gemini API for both AI detection and sentiment analysis tasks • Persistent Memory Architecture: Discovered how Backboard.io's stateful AI memory differs from traditional caching solutions and enables smarter applications • Browser Extension Development: Mastered Chrome extension architecture, content scripts, and cross-origin communication • Fraud Detection Patterns: Learned common indicators of e-commerce scams and how to algorithmically detect them using multiple AI modalities

What's next for DEMM Listing Inspector Browser Extension

Expand Platform Support: Extend beyond Etsy to Amazon, eBay, AliExpress, and other e-commerce platforms Enhanced CLIP Integration: Fine-tune CLIP on e-commerce product datasets to improve accuracy for product-specific comparisons Seller History Analysis: Track seller reputation over time and flag accounts with suspicious pattern changes Community Reporting: Allow users to report scams and build a collaborative fraud database using Backboard.io's shared memory capabilities Price Anomaly Detection: Use ML to identify listings with suspicious pricing patterns compared to similar products Browser Dashboard: Create a comprehensive dashboard showing all scanned listings and fraud trends over time Mobile App: Extend protection to mobile shoppers with iOS and Android apps Duplicate Listing Detection: Use CLIP embeddings to identify when the same AI-generated image is used across multiple fraudulent listings Key Technologies: Google Gemini (SynthID), OpenAI CLIP, Backboard.io, Chrome Extensions, Flask, Python, TextBlob Team: DEMM