An interactive image segmentation web application based on SAM 2.1 (Segment Anything Model 2). After uploading an image, users can click on any object to segment and highlight it in real-time.

Tech Stack

• Backend: FastAPI + SAM 2.1 (Meta's segmentation model, Apache 2.0 Open Source)
• Frontend: Next.js 15 + TypeScript + Tailwind CSS v4 + Framer Motion
• Model: SAM 2.1 Hiera — supports image/video segmentation; no application required, direct download
• Package Management: uv (Backend) + npm (Frontend)

Key Features

• Click-to-Segment: Interactive segmentation with millisecond-level response
• Multiple Model Selection: tiny (39M) / small (46M) / base+ (81M) / large (224M)
• CPU/GPU Dual Mode: The tiny model can even run on a CPU
• Multi-Object Selection: Annotate multiple segmented areas with different colors
• RLE Compressed Transmission: Mask data compression rate > 98%
• No Application Required: Model weights are downloaded directly under the Apache 2.0 license

Upload a PDF and ask your question to get accurate answers based on the document’s content. The system prioritizes searching local documents by default. If no answer is found in the document, it automatically falls back to web search and includes self-correction capabilities.

A lightweight personal assistant app that integrates weather notifications, reminder tasks, and AI chat capabilities. Built with PWA technology, it requires no app store—simply access it through your browser to install and use.

This project is built with Ultralytics YOLOv8 and Flask Blueprints, providing an intelligent pose estimation web application platform with the following key features:

• Multi-module Pose Estimation Supports pose estimation for images, videos, and real-time camera streams. The modular design keeps each function clear and independent.

• User Management System Includes user registration, login, and profile management, with persistent data storage powered by PostgreSQL.

• Frontend–Backend Interaction Utilizes JavaScript, jQuery, and other frontend technologies to enable image uploading, video playback, and real-time camera data exchange.

• Efficient Model Inference Integrates the YOLOv8 model to deliver fast and accurate human keypoint detection and pose estimation.

• Complete Project Architecture Covers database configuration, backend API design, frontend development, and deployment, forming a fully operational intelligent pose estimation web service.

A YOLO label analysis and integration tool built with Electron.

This project leverages a vector database to build a facial recognition–based registration and login system, featuring the following core components:

• Vector Database Support Utilizes PostgreSQL with the pgvector extension to enable efficient vector storage and similarity search, significantly improving facial feature matching performance.

• Integrated Facial Recognition Model Incorporates the InsightFace model for accurate facial feature extraction, supporting multiple distance metrics such as Euclidean distance and cosine similarity.

• Multi-Platform Support Provides a Python CLI module for registration and login, allowing quick testing and usage. A modular web application is built using Flask 3 Blueprints, enabling users to register and log in directly through the web interface.

• Database Management Uses PostgreSQL to store user vector embeddings and related metadata, with support for database initialization and extension activation.

• Automated Testing Integrates Selenium-based automated test scripts to perform end-to-end testing of the registration and login pages.