Projects

Overview: ultralyticsPro 是一个专注于YOLO系列改进模型的项目，包括YOLO11、YOLOv8、YOLOv10、RT-DETR、YOLOv7和YOLOv5的改进版本。该项目旨在通过优化backbone、neck、head、loss、IoU、NMS等模块来提升模型性能。 Key features: - 支持YOLO11改进 - 支持YOLOv8改进 - 支持YOLOv10改进 - 支持RT-DETR改进 - 支持YOLOv7改进 - 支持YOLOv5改进 Tech stack: Python Why it matters: ultralyticsPro 通过优化关键模块，显著提升了YOLO系列模型的性能，适用于各种计算机视觉应用，如目标检测、跟踪和识别等。

Overview: database.build is an in-browser Postgres sandbox with AI assistance, allowing users to instantly spin up and manage Postgres databases directly in their browser, paired with large language models for various use cases. Key features: - Drag-and-drop CSV import - Generate and export reports - Generate charts - Build database diagrams Tech stack: TypeScript Why it matters: database.build provides a powerful Postgres sandbox with AI assistance, making it ideal for developers looking to experiment with Postgres without the need for a remote container or WebSocket proxy. The ability to generate and export reports, charts, and database diagrams adds significant value to the user experience.

Overview: This lightweight WebGL and JavaScript library is designed for real-time multi-face detection, tracking, and augmented-reality face filters. It supports rotation, mouth-opening, and integrates with popular 3D engines like THREE.js, Babylon.js, and Canvas2D. Key features: - Face detection - Face tracking - Face rotation detection - Mouth opening detection - Multiple faces detection and tracking - Robust for various lighting conditions - Video acquisition with HD capabilities - Mobile-friendly - Interfaced with 3D engines like THREE.js, Babylon.js, and A-FRAME - Interfaced with accessible APIs like CANVAS, CSS3D Tech stack: JavaScript Why it matters: This library provides a powerful and flexible solution for real-time face detection and tracking, making it ideal for augmented reality applications and interactive user interfaces. The lightweight nature and compatibility with multiple engines make it suitable for a wide range of projects, from mobile devices to web applications.

Overview: DeepDanbooru is an AI-based multi-label image classification system designed for estimating tags on anime-style girl images. It leverages TensorFlow to process and classify images efficiently. Key features: - Real-time tag estimation on anime-style girl images - Compatibility with various Python versions (3.11) - Supports multiple datasets and filtering options - Easy installation and usage through a simple command-line interface Tech stack: Python 3.11, TensorFlow, Click, numpy, requests, scikit-image, six, tensorflow-io Why it matters: DeepDanbooru simplifies the process of tagging anime-style girl images, making it easier for artists and enthusiasts to identify and categorize their work. The system's ability to handle multiple tags and optional filtering options makes it highly versatile and useful for various applications in the anime community.

### Overview: SimpleTuner is a general fine-tuning kit designed for image, video, and audio diffusion models, aimed at making the process of training and tuning models as simple as possible. ### Key Features: - **User-friendly web UI**: Manages the entire training lifecycle through a sleek dashboard. - **Multi-modal training**: Unified pipeline for Image, Video, and Audio generative models. - **Multi-GPU training**: Distributed training across multiple GPUs with automatic optimization. - **Advanced caching**: Image, video, audio, and caption embeddings cached to disk for faster training. - **CaptionFlow integration**: Generates dataset captions using [bghira/CaptionFlow](https://github.com/bghira/CaptionFlow). - **Aspect bucketing**: Support for varied image/video sizes and aspect ratios. - **Concept sliders**: Slider-friendly targeting for LoRA/LyCORIS/full with positive/negative/neutral sampling. - **Memory optimization**: Most models trainable on 24G GPU, many on 16G with optimizations. - **DeepSpeed & FSDP2 integration**: Train large models on smaller GPUs with optim/grad/parameter sharding, context parallel attention, gradient accumulation. ### Tech Stack: Python ### Why it Matters: SimpleTuner aims to simplify the process of training and tuning diffusion models, making it accessible and efficient for a wide range of applications.

### Overview: This repository contains PyTorch implementations of various Deep NLP models, focusing on NLP with Deep Learning, as covered in the Stanford course CS224N. The models include Skip-gram, Negative Sampling, GloVe, Window Classifier for Named Entity Recognition, Neural Dependency Parser, RNN Language Model, Neural Machine Translation, CNN for Text Classification, Recursive Neural Network for Sentiment Classification, and Dynamic Memory Network for Question Answering. These models are designed to help researchers and developers deepen their understanding and skills in natural language processing. ### Key Features: - **Skip-gram-Naive-Softmax** - Notebook: [notebook](https://nbviewer.jupyter.org/github/DSKSD/DeepNLP-models-Pytorch/blob/master/notebooks/01.Skip-gram-Naive-Softmax.ipynb) - Paper: [arxiv.org/abs/1301.3781](https://arxiv.org/abs/1301.3781) - **Skip-gram-Negative-Sampling** - Notebook: [notebook](https://nbviewer.jupyter.org/github/DSKSD/DeepNLP-models-Pytorch/blob/master/notebooks/02.Skip-gram-Negative-Sampling.ipynb) - Paper: [nips.cc/Paper/5021-distributed-representations-of-words-and-phrases-and-their-compositionality.pdf](http://papers.nips.cc/paper/502

Overview: Stop That Shit is a project designed to protect AI coding agents in Codex/GPT workflows by intercepting unrequested hashes, checksums, and task-scope creep. It aims to streamline the workflow by ensuring only necessary actions are performed. Key features: - Intercept unrequested hashes, checksums, and task-scope creep - Simplify the workflow by removing unnecessary checks and validations - Ensure that only the necessary actions are performed Tech stack: JavaScript Why it matters: This project helps streamline AI coding workflows by preventing unnecessary actions and ensuring that only the necessary steps are taken, thereby improving efficiency and reducing errors. Repository: lennney/stop-that-shit Primary language: JavaScript GitHub stars: 2938 GitHub description: Stop That Shit（别再造史了）｜面向 Codex/GPT 场景的多平台 Hook + Skill Guard：拦截 AI coding agent 无需求的哈希、校验和与任务范围膨胀。 A multi-platform Hook + Skill Guard for AI coding agents in Codex/GPT workflows: stop unrequested hashes, checksums, and task-scope creep. README excerpt: <p align="center"> <img src="assets/stop-stamp.svg" alt="Stop That Shit（别再造史了）AI Agent 任务边界 Guard 的红色 STOP 印章" width="240"> </p> <h1 align="center">Stop That Shit（别再造史了）</h1> <p align="center"> <a href="https

### Overview: Memobase is a user profile-based memory system designed to enhance long-term user memory in AI chatbot applications. It aims to bring personalization, understanding, and continuous learning to your AI agents. ### Key Features: - **User Profile-Based Memory**: Empowers your AI to remember, understand, and evolve with your users. - **Performance**: Achieves top-tier search performance in the LOCOMO benchmark. - **LLM Cost**: Includes a buffer for each user to efficiently distribute chat processing overhead. - **Latency**: Provides access to user profiles and timelines without pre-processing, keeping online latency under 100ms. ### Tech Stack: Python ### Why It Matters: Memobase offers a perfect balance for various memory solutions, enhancing the performance, cost efficiency, and latency of your AI applications. It ensures seamless integration and optimal user experience.

### Overview: ClawPanel 是一个支持多 AI Agent 框架的可视化管理面板，内置智能 AI 助手，提供一键安装、自动诊断配置、排查问题、修复错误等功能。它支持 OpenClaw、Hermes Agent、DeepSeek Harness、OpenCode 和 Pi 等多种 AI 管理工具。 ### Key features: - **内置智能 AI 助手**：一键安装、自动诊断配置、排查问题、修复错误。 - **8 大工具 + 4 种模式 + 交互式问答**：从新手到老手都能轻松管理。 - **运行时管理**：支持固定验收版本、独立工作台、统一模型渠道、Web/headless 安全内嵌。 - **Hermes Agent**：会话、记忆、人格与工具全景管理。 - **DeepSeek Harness**：面板内完成配置与实际操作。 ### Tech stack: JavaScript ### Why it matters: ClawPanel 是一个多功能的多引擎 AI 管理面板，适用于多种 AI 工具和应用场景，提供高效、便捷的管理体验。

Overview: MiniAGI is a simple general-purpose AI agent built on the OpenAI API, designed for tasks such as programming, web searches, and creative tasks. Key features: - Writing Python code to draw a car - Creating a website and serving it locally - Generating original online job ideas Tech stack: Python Why it matters: MiniAGI simplifies complex tasks by providing a user-friendly interface and minimal setup, making it accessible for both beginners and advanced users. Repository: muellerberndt/mini-agi Primary language: Python GitHub stars: 2933

### Overview: Handtrack.js is a library designed for real-time hand detection in the browser, enabling developers to prototype and integrate hand tracking directly into their projects. ### Key features: - Real-time hand bounding box detection - Supports multiple hand classes (open, closed, pinch, point, face) - Updated dataset for improved accuracy and functionality - New classes and labels for enhanced user experience - Optimized model sizes for faster inference ### Tech stack: JavaScript ### Why it matters: Handtrack.js simplifies the process of integrating hand tracking into web applications, making it easier for developers to create interactive and engaging user experiences without the need for complex backend setups.

Overview: EasyEdit is an easy-to-use knowledge editing framework designed for Large Language Models (LLMs) at the Conference on Artificial Intelligence (ACL 2024). It simplifies the process of controlling and modifying the knowledge and behavior of LLMs during inference. Key features: - Real-time steering of LLMs during inference - Unifying framework for controllability without retraining - Knowledge insert, update, and erase functionalities Tech stack: Jupyter Notebook Why it matters: EasyEdit provides a user-friendly interface for controlling LLMs, making it accessible to researchers and practitioners without requiring extensive knowledge of LLM internals. This framework enables researchers to experiment with different steering methods and evaluate their effectiveness in real-time, without the need for retraining the model. Repository: zjunlp/EasyEdit GitHub stars: 2930 GitHub description: [ACL 2024] An Easy-to-use Knowledge Editing Framework for LLMs.

Overview: NLP Architect is an open-source Python library designed for exploring state-of-the-art deep learning topologies and techniques in Natural Language Processing and Natural Language Understanding neural networks. Key features: - Core NLP models - Novel NLU models - Optimized NLP/NLU models - Model-oriented design: command-line training, Python API, custom processes, CLI sub-system - Frameworks: TensorFlow, PyTorch, Dynet - Essential utilities: text preprocessing, IO, data manipulation, metrics, embeddings Tech stack: Python Why it matters: NLP Architect provides a comprehensive and flexible solution for optimizing and showcasing natural language processing and understanding neural networks, making it valuable for researchers and developers alike.

### Overview: LightX2V is a lightweight image/video generation inference framework designed to efficiently and performantly generate images and videos. It integrates advanced techniques for various tasks such as text-to-video (T2V), image-to-video (I2V), text-to-image (T2I), and image editing (I2I). ### Key Features: - **Text-to-Video (T2V)**: Converts text inputs into video outputs. - **Image-to-Video (I2V)**: Converts images into videos. - **Text-to-Image (T2I)**: Converts text inputs into images. - **Image Editing (I2I)**: Edits images to enhance or modify their appearance. ### Tech Stack: - Python ### Why It Matters: LightX2V offers a powerful and efficient solution for generating images and videos, making it ideal for applications requiring real-time or near-real-time processing. Its lightweight nature makes it suitable for both cloud and local deployment, ensuring fast and reliable performance. ### Repository: Repository: ModelTC/LightX2V GitHub Stars: 2921 ### Primary Language: Primary Language: Python ### GitHub Description: LightX2V is an advanced lightweight image/video generation inference framework engineered to deliver efficient, high-performance image/video synthesis solutions. This unified platform integrates multiple state-of-the-art image/video generation techniques, supporting diverse generation tasks including text-to-video (T2V), image-to-video (I2V), text-to-image (T2I), image-editing (I2I). ### Documentation &

### Overview: InternLM-XComposer2.5-OmniLive is a comprehensive multimodal system designed for long-term streaming video and audio interactions. It leverages advanced techniques to enhance the understanding and interaction capabilities of video and audio content. ### Key Features: - **Multimodal Interaction**: Integrates video and audio streams for real-time, interactive analysis. - **Long-Term Streaming Support**: Capable of processing and analyzing video and audio data over extended periods. - **Versatile Training Data**: Utilizes a diverse set of training data to improve model performance. ### Tech Stack: - Python ### Why It Matters: This system is crucial for applications requiring real-time analysis and interaction with video and audio streams, such as remote support, surveillance, and educational tools. Its ability to handle long-term data streams makes it highly valuable in various industrial and educational settings. ### Repository: InternLM/InternLM-XComposer ### GitHub Stars: 2,928 ### GitHub Description: InternLM-XComposer2.5-OmniLive: A Comprehensive Multimodal System for Long-term Streaming Video and Audio Interactions ### Tech Stack: Python ### Why It Matters: Enhances real-time analysis and interaction capabilities with video and audio streams, making it valuable in various industrial and educational applications.

Overview: This repository contains a PyTorch implementation of the BigGAN model, authored by Andrew Brock and Alex Andonian. It includes code for training BigGANs on the ImageNet dataset and provides various training scripts and metrics for evaluation. Key features: - Training of BigGAN models using PyTorch - Support for gradient accumulation for faster training - Preprocessing of the ImageNet dataset for faster I/O - Generation of samples and interpolations for evaluation - Metrics for evaluation: Inception Score (IS), Frechet Inception Distance (FID) Tech stack: Python Why it matters: This implementation provides a useful tool for training and evaluating BigGAN models, which are known for their ability to generate high-fidelity images from noise. Repository: ajbrock/BigGAN-PyTorch GitHub stars: 2928

### Overview: JADX-AI-MCP is a fully automated MCP server plugin designed to integrate with the Zin MCP Suite. It leverages the powerful capabilities of Jadx to analyze Android APKs using Large Language Models like Claude, enabling the detection of vulnerabilities, APK analysis, and reverse engineering. ### Key Features: - **Fully Automated MCP Server Integration**: Seamless integration with the MCP Suite for efficient server management. - **Jadx Plugin**: Utilizes Jadx to analyze APKs, providing detailed insights and reverse engineering capabilities. - **LLM Integration**: Communicates with Large Language Models (e.g., Claude) to enhance the analysis process. ### Tech Stack: - Java 11+ - Python 3.10+ ### Why It Matters: JADX-AI-MCP simplifies the process of analyzing Android APKs, making it easier for developers to identify vulnerabilities and improve security measures. Its integration with the MCP Suite ensures a comprehensive approach to software security and reverse engineering.

Overview: KerasTuner is a lightweight, easy-to-use hyperparameter tuning library designed specifically for Keras, addressing common pain points in hyperparameter search. Key features: - Define-by-run syntax for flexible hyperparameter configuration - Built-in support for Bayesian Optimization, Hyperband, and Random Search algorithms - Scalable and easy to extend for researchers to experiment with new search algorithms Tech stack: Python Why it matters: KerasTuner simplifies the process of hyperparameter tuning, making it accessible to both beginners and advanced researchers, while also providing a robust framework for experimenting with various optimization techniques.

### Overview: Claude Code learns from your corrections, but context is lost. Pro Workflow solves this by storing corrections, research wikis, and skills in a single SQLite database, ensuring persistent learning and context compaction. ### Key Features: - **Self-Correction Memory**: Every correction becomes a rule, searchable in FTS5-indexed wikis. - **Persistent Research Wikis**: Research wikis grow overnight, with auto-research loop for continuous updates. - **Quality Gates**: LLM-powered hooks, deterministic git/secret guards, and compaction-aware state. - **41 Skills, 8 Agents, 23 Commands**: Works seamlessly with Claude Code, Cursor, and 32+ agents via skills add. ### Tech Stack: JavaScript ### Why It Matters: Pro Workflow ensures Claude Code learns and grows continuously, without losing context or repeating mistakes.

Overview: forkd is a lightweight AI agent microVM sandbox runtime that supports rapid spawning and branching of child processes, providing KVM isolation and efficient memory sharing. Key features: - Forks 100 microVMs in 101 ms - Branches a live VM in 56 ms - Supports live branching with ~56 ms p50 / 64 ms p90 pause time - Built on Firecracker and compatible with Python, Java, and ML models Tech stack: Rust Why it matters: forkd simplifies AI agent deployment and testing by providing a lightweight, isolated environment for rapid development and testing, while maintaining high performance and efficiency.