Background

I'm a fashion enthusiast who shops on Grailed, a secondhand clothing marketplace. The platform has a search problem: you can only find items if you know the brand name or item name. For vintage pieces without clear labels, you're stuck. Even worse, if you like a piece but the condition or price isn't right, there's no way to find similar styles.

Grailed listings aren't standardized—the same item can look completely different depending on angles, lighting, and backgrounds. I wanted to search by what actually matters: visuals. Patterns, colors, silhouettes, details.

Solution

What I Built

A full-stack visual search engine with four components:

Backend API — Scrapes listings, generates CLIP embeddings, handles similarity search
Admin Dashboard — Manages scraping and embedding processes
User Frontend — Displays search results
Chrome Extension — Main entry point; click on any Grailed listing to find similar items

How It Works

I scraped ~2 million listings from Grailed (images, titles, metadata). Each image and title gets processed through OpenAI's CLIP model to generate embeddings—vector representations that capture visual and semantic meaning. Embeddings are stored in Pinecone for fast similarity search.

When you open the Chrome extension on a Grailed listing, it sends the image to my backend, searches the vector database, and returns visually similar items in the frontend.

The Challenge

This wasn't a single app—it was a complete data pipeline. The hardest part was deciding the tech stack and architecture:

Scraping 2M listings with proper concurrency and checkpointing
Cleaning and organizing massive datasets in MongoDB
Batch processing images through CLIP efficiently
Setting up Pinecone with the right indexing strategy
Making the flow seamless (extension → backend → frontend)

I've discussed this project in multiple interviews because it touches on full-stack development, ML systems, and working with data at scale.

Tech Stack

Backend

Python 3.12, Quart (async FastAPI) — API server
PyTorch + Transformers — CLIP model for embeddings
Playwright — Web scraping
MongoDB — Raw data storage
Pinecone — Vector similarity search
Docker, Google Cloud Secret Manager — Deployment

Frontend

Next.js 13+, React, Tailwind CSS — Dashboard and user interface

Browser Extension

Chrome Extension API, JavaScript — Grailed integration

Pipeline

2M listings scraped and processed
512-dimensional CLIP embeddings for images and text
Checkpoint system for resumable long-running processes
RESTful API with Server-Sent Events for real-time logs

MY WORK @ BAIDU

SELECT PROJECT

®AYMOND