Projects

### Overview: AI Maestro is an AI Agent Orchestrator that enables seamless communication and collaboration among multiple AI agents, allowing for efficient management and orchestration of these agents across different machines and locations. ### Key Features: - **Persistent Memory:** Agents retain their context and history across sessions. - **Agent-to-Agent Messaging:** Direct communication between agents for seamless collaboration. - **Multi-Machine Support:** Orchestrates agents across multiple computers without the need for a central server. - **Orchestration Dashboard:** Centralized dashboard to manage and monitor all agents in one place. ### Tech Stack: TypeScript ### Why It Matters: AI Maestro simplifies the management and collaboration of AI agents, enabling organizations to harness the power of AI more effectively and efficiently.

Overview: 4KAgent is an intelligent computer vision agent designed to restore any image to 4K resolution, addressing various image categories and degradation levels. Key features: - Universal any-image-to-4K upscaling - Large vision-language model-based restoration - Quality-driven Mixture-of-Experts policy - Face restoration pipeline - Profile module for customization Tech stack: Python Why it matters: 4KAgent offers a versatile solution for image restoration, addressing a wide range of image categories and degradation levels, making it a valuable tool in various applications such as remote sensing, microscopy, and biomedical imaging.

### Overview: Cezar is an open-source orchestrator designed for running Claude Code, Codex, OpenCode, Pi, and other AI coding agents in parallel. It allows you to run these agents locally or on a VPS, ensuring they continue working even when you're away. ### Key features: - 🚀 Free and open-source - 🖥️ Uses your own Claude, Codex, Copilot, Cursor, Junie, Opencode, or Pi login without API keys - ☁️ Easy to set up on a VPS, keeping agents running when your laptop is closed - 📱 Fully responsive, starting and reviewing tasks from your phone - 🔀 Every task gets its own git worktree, allowing multiple agents to work simultaneously - 🤖 Turn on Autonomous mode and let the run never stop, finishing automatically - 📡 Watch it work live: agent text, tool calls, tokens, and cost - 🏁 Run the same task 2 or 3 times, compare diffs, and keep the best one - 🧩 Skills are Markdown files, workflows are short YAML files, and agents can be mixed per step - 🐙 Run the agent directly on a GitHub issue, nothing merges on its own - 📂 One cockpit for all your projects - 📊 Workspace dashboard shows what needs, what is running, and what finished across projects - 💸 Track reported cost and token usage by project in Usage & cost - ⏰ Schedule recurring work or launch tasks from

### Overview: Halo is an open-source framework designed by White Circle for training large language and multimodal models. It supports various training methods and parallelism techniques, making it highly efficient and scalable. ### Key Features: - **Parallelism Techniques:** EP, CP, TP, ETP, FSDP2, BF16 training, distributed rollouts. - **Training Methods:** Pre-training, SFT, preference optimization, fully asynchronous multi-turn RL. - **Model Integration:** Trainers subclass Hugging Face or TRL trainers, checkpoints remain standard SafeTensors, and `from_pretrained` continues to work. - **Performance Optimizations:** DeepEP V2, FlashAttention 4, Liger, Grouped GEMM, padding-free packing, AdamWBF16, and more. ### Tech Stack: Python ### Why It Matters: - **Direct Training:** Trainers subclass Hugging Face or TRL trainers, checkpoints remain standard SafeTensors, and `from_pretrained` continues to work. - **Parallelism:** EP wraps MoE blocks, CP wraps attention, and TP/ETP shard weights in place. There is no separate distributed implementation of the model. - **Distributed Stack:** FSDP2, DTensor, and DeviceMesh provide the underlying primitives, while DeepEP handles the all-to-all communication for Expert Parallelism. EP, CP, TP, and ETP each have their own setting; EP combines with CP, TP, or ETP across GPUs and nodes, and other pairings are rejected at config time. -

### Overview: EvolvingLMMs-Lab/lmms-engine is a simple, unified multimodal models training engine designed for efficient and flexible model training at scale. It is lean, flexible, and built for rapid experimentation and hacking. ### Key Features: - **Unified Model Support**: Supports a wide range of multimodal models. - **Lean Architecture**: Designed for efficient use of resources. - **Flexible Configuration**: Easily configurable for various training setups. - **Optimizations**: Includes performance optimizations like FlashAttention and LigerKernel for better efficiency. ### Tech Stack: - Python 3.11+ - PyTorch - uv package manager ### Why It Matters: - **Efficiency**: Provides comprehensive Model FLOPs Utilization (MFU) metrics for benchmarking various architectures and configurations. - **Unified Training**: Supports a unified framework for training multimodal models, making it easier to experiment with different architectures and configurations. ### GitHub Stars: 832 ### Repository: - **Repository**: https://github.com/EvolvingLMMs-Lab/lmms-engine - **Primary Language**: Python - **GitHub Stars**: 832 ### Quick Start: - **Clone Repository**: `git clone https://github.com/EvolvingLMMs-Lab/lmms-engine.git` - **Install Dependencies**: `uv pip install -e .[all]` or `uv pip install -e .` - **Optional: Performance Optimizations**: `uv pip install flash-attn` and `uv pip install liger-kernel

Overview: AgnosticUI Local (v2) is a CLI-based UI component library designed to be integrated seamlessly into various frameworks, providing high-fidelity context and AI-readability. Key features: - Copies components directly into your project - Works with AI tools and agent-driven UIs - Framework-agnostic, built for humans and AI agents Tech stack: TypeScript Why it matters: AgnosticUI Local simplifies the development process by providing a context-aware, framework-agnostic, and AI-ready solution, making it easier to create and maintain complex, accessible workflows using AI-driven interfaces.

Overview: pg_vectorize is a Postgres extension and HTTP server that automates the transformation and orchestration of text to embeddings, providing full-text, semantic, and hybrid search capabilities. This project simplifies the integration of vector search, full text search, and hybrid search into any Postgres database, enabling the development of RAG and search engines. Key features: - An external server implementation for managed Postgres databases - A Postgres extension for in-database integration - Provides hooks into popular LLMs for seamless integration with large language models Tech stack: Rust Why it matters: pg_vectorize simplifies the process of integrating vector search, full text search, and hybrid search into Postgres, making it easier to build RAG and search engines on the PostgreSQL platform.

### Overview: GaussianDreamer is a fast and efficient framework for generating 3D objects from text prompts, bridging the strengths of 2D and 3D diffusion models. It achieves high-quality 3D generation within 15 minutes on a single GPU, making it significantly faster than existing methods. ### Key Features: - **Fast Generation**: Generates 3D instances or avatars within 15 minutes on one GPU. - **High-Quality Output**: Achieves high-quality 3D generation, ensuring consistency and realism. - **Integration with Unity**: Supports seamless integration into Unity game engines, enabling real-time rendering. - **Versatile Applications**: Can be used for 3D asset generation, avatar creation, and more. ### Tech Stack: - Python ### Why It Matters: - **Efficiency**: Outperforms previous methods by generating high-quality 3D objects faster. - **Versatility**: Integrates seamlessly with Unity, making it useful for game development. - **Real-Time Rendering**: Supports real-time rendering, enhancing the user experience. ### Repository: - GitHub Repository: [hustvl/GaussianDreamer](https://github.com/hustvl/GaussianDreamer) - GitHub Stars: 832 - GitHub Description: [CVPR 2024] GaussianDreamer: Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models ### Updates: - 1/13/2025: Released GaussianDreamerPro with enhanced quality.

### Overview: This repository is a collection of papers and resources focused on data augmentation techniques for Natural Language Processing (NLP). It serves as a comprehensive guide to the various methods used to enhance the quality and diversity of training data for NLP models. ### Key Features: - **Text Classification**: Papers on unsupervised word sense disambiguation, synonym replacement, and lexical and frame-semantic embedding based data augmentation. - **Translation**: Techniques for improving translation quality through data augmentation. - **Summarization**: Methods for generating summaries that are more informative and contextually relevant. - **Question-Answering**: Approaches to enhance the accuracy and relevance of question-answering models. - **Sequence Tagging**: Techniques for improving the performance of sequence tagging models. - **Parsing**: Methods for improving the accuracy of parsing models. - **Grammatical Error Correction**: Tools for correcting grammatical errors in text. - **Generation**: Techniques for generating text that is more diverse and contextually relevant. - **Dialogue**: Approaches for improving dialogue systems. - **Multimodal**: Methods for combining text and other modalities to improve NLP models. - **Mitigating Bias**: Techniques to reduce bias in training data. - **Mitigating Class Imbalance**: Methods to balance the representation of different classes in training data. - **Adversarial Examples**: Techniques to improve the robustness of NLP models against adversarial attacks. - **Compositionality**: Approaches to improve the compositional abilities of NLP models. - **Automated Augmentation**: Tools and

### Overview: This repository provides the official implementation of "RIFLEx: A Free Lunch for Length Extrapolation in Video Diffusion Transformers" (ICML 2025), "UltraViCo" (ICLR 2026), and "UltraImage" (ICLR 2026), enabling diffusion-transformer extrapolation for long video generation and high-resolution image generation in a plug-and-play manner. ### Key features: - **RIFLEx**: Supports HunyuanVideo-diffusers and CogVideoX-diffusers. - **UltraViCo**: Offers UltraViCo for Wan2.1 and HunyuanVideo for HunyuanVideo. - **UltraImage**: Supports UltraImage for Flux and UltraImage for Qwen-Image. ### Tech stack: Python ### Why it matters: These models significantly advance the field of video generation by enabling efficient and accurate extrapolation of video sequences, making them invaluable tools for researchers and practitioners in the field of computer vision and artificial intelligence.

Overview: GraphGPT is a research project that focuses on graph instruction tuning for large language models. It aims to enhance the performance of LLMs by leveraging graph-based instruction tuning techniques. Key features: - Efficient and lightweight training code for large language models. - Two-stage instruction tuning on multiple GPUs for improved performance. - Comprehensive documentation and resources for both project and demo pages. Tech stack: Python Why it matters: GraphGPT offers a practical approach to enhance the capabilities of large language models by utilizing graph-based instruction tuning, which can lead to significant improvements in performance and efficiency. Repository: GitHub - HKUDS/GraphGPT GitHub stars: 831 GitHub description: "GraphGPT: Graph Instruction Tuning for Large Language Models"

Overview: MeshDiffusion is an official implementation of a score-based generative model for 3D mesh generation. This model is designed to generate high-quality 3D meshes with a direct parametrization of deep marching tetrahedra (DMTet). Key features: - Direct parametrization of deep marching tetrahedra (DMTet) - Pretrained models for various 3D objects - Datasets with processed cubic grids for deformation scale and SDF values - Inference capabilities for unconditional and conditional generation Tech stack: Python Why it matters: MeshDiffusion offers a novel approach to 3D mesh generation, providing high-quality results with a direct parametrization method. This model is particularly useful for applications requiring precise and efficient 3D mesh modeling.

### Overview: lightNLP 是一个基于 PyTorch 和 torchtext 的自然语言处理深度学习框架，旨在提供基础的自然语言处理任务实现。 ### Key features: - 命名实体识别 (NER) - 词性标注 (POS) - 依存分析 (Dependency Parsing) - 情感分析 - 文本蕴含 (Semantic Similarity) - 文本生成 (Text Generation) - 语言模型 (Language Model) - 词性标注 (POS) - 事件抽取 (Event Extraction) - 自动摘要 (Automatic Summarization) - 机器翻译 (Machine Translation) - 句子相似度 (Sentence Similarity) - 序列到序列 (Seq2Seq) - 聊天机器人 (Chatbot) - 命名实体识别 (NER) - 句子转换 (Sentence Conversion) ### Tech stack: Python ### Why it matters: 本项目为初学者和对自然语言处理感兴趣的开发者提供了一个基础的框架，旨在帮助他们快速上手并实践自然语言处理技术。

Overview: WilmerAI is an advanced semantic prompt router designed to enhance the functionality of applications that integrate with multiple LLMs via REST APIs. It uses a node-based workflow engine to route prompts across different LLMs, ensuring that the context of the conversation is fully understood. Key features: - Multi-layer prompt routing - Complex workflows for integrating with LLMs - Node-based workflow engine for efficient task orchestration Tech stack: Python Why it matters: WilmerAI simplifies the integration of LLMs into applications, allowing for more context-aware and versatile chatbot creation. Its ability to handle multiple LLMs and workflows makes it a valuable tool for developers looking to enhance their AI applications. Repository: SomeOddCodeGuy/WilmerAI Primary language: Python GitHub stars: 830 GitHub description: WilmerAI is an application designed for advanced semantic prompt routing and complex task orchestration. It originates from the need for a router that could understand the full context of a conversation, rather than just the most recent message. Unlike simple routers, WilmerAI's routing system can analyze the entire conversation history, enabling it to understand the true intent behind a query. This contextual understanding is made possible by its core: a node-based workflow engine, which is responsible for orchestrating different LLMs, external tools, and custom scripts. To the client application, this entire multi-step process appears as a standard API call, enabling advanced backend logic without requiring changes to your existing front-end tools. WilmerAI is designed for developers looking to enhance their AI

Overview: This repository contains complete application examples developed using Firecrawl, showcasing various implementations and use cases. Key features: - Websites - Projects - Other applications Tech stack: Jupyter Notebook Why it matters: These examples provide practical demonstrations of how Firecrawl can be applied to different projects, making it easier for developers to understand and implement the technology effectively.

Overview: LongMem is an official implementation of the NeurIPS 2023 paper "Augmenting Language Models with Long-Term Memory". It provides a comprehensive framework for memory-augmented language models. Key features: - Memory-Augmented Adaptation Training - Pre-trained Language Models - Transformer Decoder with SideNetwork - Memory Bank and Retrieval - Joint Attention for Memory Fusion Tech stack: Python Why it matters: LongMem offers a robust solution for enhancing language models by incorporating long-term memory, making it a valuable addition to the field of language understanding and artificial intelligence.

Overview: Lingua is a highly accurate natural language detection library for Java and the JVM, designed to work with both long and short text, including single words and phrases. It offers a simple, rule-based and statistical approach without the need for external APIs or services, making it suitable for preprocessing linguistic data in applications like text classification and spell checking. Key features: - High accuracy in detecting the language of text - Supports 75 languages, including English, Spanish, and German - Works with both short and long texts, including single words and phrases - No configuration required, making it easy to use offline Tech stack: Kotlin Why it matters: Lingua simplifies language detection by providing accurate results on short and long texts, eliminating the need for lengthy text fragments and complex configuration. This makes it a valuable tool for applications requiring linguistic preprocessing, such as text classification and spell checking, without the complexity and potential inaccuracies of more comprehensive libraries.

### Overview: Taichi NeRFs is a PyTorch + Taichi implementation of Instant NGP NeRF training pipelines. This project provides comprehensive training scripts for both synthetic NeRF datasets and real-world video datasets, along with a detailed blog post on modeling techniques. ### Key Features: - **Synthetic NeRF**: Training and visualization of a synthetic NeRF dataset using Taichi and PyTorch. - **360_v2 Dataset**: Training a NeRF model on a real-world dataset using Taichi and PyTorch. - **Real-World Video**: Training a NeRF model using a real-world video dataset, with adjustable parameters for better performance. ### Tech Stack: - Taichi - PyTorch - Colmap ### Why It Matters: This project offers a practical solution for training NeRF models using modern graphics hardware, making it accessible for both research and industrial applications. The use of Taichi and PyTorch ensures efficient performance and seamless integration with existing machine learning workflows. ### Repository: - Repository: taichi-dev/taichi-nerfs - GitHub Stars: 829 - GitHub Description: Implementations of NeRF variants based on Taichi + PyTorch ### Installation: - **Prerequisites**: - Install PyTorch: `python -m pip install torch torchvision --extra-index-url https://download.pytorch.org/whl/cu116` - Install Taichi: `pip install -U pip && pip install -i https://pypi.taichi.graphics/simple/ taichi-night

Overview: ChatGenTitle 是一个使用百万 arXiv 论文信息在 LLaMA 模型上进行微调的论文题目生成模型。它能够为科研人员提供高质量的论文题目建议。 Key features: - 使用百万 arXiv 论文信息进行微调 - 在 LLaMA 模型上进行微调 - 提供论文标题生成功能 Tech stack: Python Why it matters: ChatGenTitle 是一个强大的工具，可以帮助科研人员快速生成高质量的论文题目，提高研究效率。

### Overview: idea-reality-mcp is an automated tool designed to check if a specific AI coding agent's startup idea already exists on various platforms. It scans GitHub, Hacker News, Stack Overflow, PyPI, and NPM to provide a 0-100 reality score with evidence, trend detection, and pivot suggestions. This helps the AI agent decide whether to build, pivot, or kill the idea before any code is written. ### Key Features: - **Reality Check**: Scans multiple databases to determine the existence and competitiveness of a startup idea. - **Score & Trend**: Provides a 0-100 score with trend direction, top competitors, and pivot suggestions. - **Market Insights**: Offers insights into market momentum and growth rates. - **Versatile Use Cases**: Ideal for startups considering new projects, ensuring they don't duplicate efforts and identify potential market gaps. ### Tech Stack: - Python ### Why It Matters: - **Market Validation**: Helps identify if similar tools already exist, ensuring the startup idea is not duplicated. - **Competitive Analysis**: Provides insights into the market's growth and trends, allowing for informed decision-making. - **Pivot Suggestions**: Offers AI-generated pivot suggestions to explore new market opportunities. ### Repository: - **Repository**: mnemox-ai/idea-reality-mcp - **GitHub Stars**: 828 - **Primary Language**: Python ### Project Status: - **Project Status (August 2026)**: Maintenance mode. The tool works, stays free & open source, and the hosted