Thrift Smart

The AI assistant that makes thrifting easier and cheaper!

Comment

Inspiration

When shopping at thrift stores or online secondhand markets, it's nearly impossible to know if you're getting a good deal. Sellers often overprice items, and buyers lack the market knowledge to negotiate confidently. I wanted to create an AI shopping assistant that could instantly analyze any product and tell you whether it's worth buying.

What it does

ThriftSmart AI analyzes product videos in real-time to determine if items are good deals. Users upload a video of any product and enter the asking price. The AI identifies the product using computer vision, searches eBay for comparable listings, compares prices, and delivers a clear verdict: COP (worth buying) or DROP (overpriced). The analysis is delivered both as text and through AI-generated voice explanation.

How I built it

We built the frontend with HTML, CSS, and JavaScript, creating a modern interface with drag-and-drop file upload. The backend runs on Flask and uses LangChain to orchestrate an AI agent. Google Gemini 3.0 Pro analyzes video content to identify products. I implemented a custom Selenium-based tool that searches eBay in headless mode and scrapes live pricing data using BeautifulSoup. The agent compares market prices against the user's asking price, generates reasoning, and ElevenLabs converts the analysis into natural speech.

Challenges I ran into

Video processing was difficult because files needed to be encoded to base64 for the Gemini API, which increased payload sizes significantly. Web scraping eBay required headless Chrome because the page is JavaScript-rendered, adding complexity with ChromeDriver management. Extracted price data was messy and inconsistent, requiring normalization logic. I also faced CORS issues during frontend-backend communication and had to implement proper preflight handling. Browser auto-play restrictions meant we needed fallback manual controls for audio playback.

Accomplishments that I'm proud of

I successfully built a complete AI agent that autonomously performs multi-step workflows: analyzing videos, conducting web research, making decisions, and explaining recommendations. The system processes everything in under 60 seconds despite the complexity. I integrated cutting-edge multimodal AI with practical web scraping to create something genuinely useful. The voice explanation adds personality and makes technical analysis accessible to everyone.

What I learned

Building AI agents that take action is far more complex than chatbots. It requires extensive error handling, state management, and tool orchestration. I learned that multimodal AI opens possibilities impossible with text alone. Real-time web scraping is challenging due to dynamic content and rate limits. I discovered that context matters for AI decisions - the same price difference means different things for different products. Finally, adding voice synthesis transformed our tool from a calculator into a trusted shopping companion.

What's next for ThriftSmart AI

I plan to add image recognition for faster analysis, expand to multiple marketplaces beyond eBay including Facebook Marketplace and Craigslist, implement historical price tracking to show trends, add barcode scanning for instant identification, build native mobile apps with camera integration, and develop condition assessment using computer vision to evaluate wear and damage. We also want to add price alerts and community features where users can share finds and verify assessments.

Share this project:

Updates