Projects

Overview: UER-py is an open-source pre-training model framework built on PyTorch, designed to facilitate the development of universal encoder representations for various natural language processing tasks. It supports pre-training on general-domain corpora and fine-tuning on downstream tasks, making it a versatile tool for research and development. Key features: - Reproducibility: UER-py has been tested on multiple datasets to ensure compatibility with original pre-training model implementations. - Model modularity: The framework is divided into components like embedding, encoder, target embedding, decoder, and target, allowing users to combine modules for flexible model construction. - Model training: It supports CPU, single GPU, and distributed training modes. - Model zoo: Users can access pre-trained models of various properties through the model zoo, which is crucial for the performance of downstream tasks. - SOTA results: UER-py supports a wide range of downstream tasks and has winning solutions in many NLP competitions. - Abundant functions: It provides functions related to pre-training, such as feature extraction and text generation. Tech stack: UER-py is built on PyTorch, with additional dependencies for TensorFlow conversion, sentencepiece tokenization, LightGBM stacking model development, CRF sequence labeling, and word segmentation tools. Why it matters: UER-py simplifies the process of pre-training and fine-tuning models, making it an essential tool for researchers and developers in the NLP field. It supports various downstream tasks and provides a comprehensive solution for pre-trained models, enhancing the performance of natural language processing applications

Overview: Distributed LLM inference allows connecting home devices into a cluster to speed up language model inference. Key features: - Supports Linux, macOS, and Windows - Optimized for ARM and x86_64 CPUs - Leverages tensor parallelism and high-speed Ethernet synchronization Tech stack: C++ Why it matters: This project enables efficient language model inference by utilizing distributed computing, making it suitable for applications requiring fast processing times. Repository: b4rtaz/distributed-llama GitHub stars: 3112 GitHub description: Distributed LLM inference. Connect home devices into a powerful cluster to accelerate LLM inference. More devices mean faster performance, leveraging tensor parallelism and high-speed synchronization over Ethernet.

Overview: This repository is a comprehensive collection of recent papers on deep learning for graphs, covering various topics such as unsupervised node representation learning, graph embedding, and adversarial network embedding. Key features: - DeepWalk: Online Learning of Social Representations - LINE: Large-scale Information Network Embedding - GraRep: Learning Graph Representations with Global Structural Information - node2vec: Scalable Feature Learning for Networks - Variational Graph Auto-Encoders - Scalable Graph Embedding for Asymmetric Proximity - Fast Network Embedding Enhancement via High Order Proximity Approximation - struc2vec: Learning Node Representations from Structural Identity - Poincaré Embeddings for Learning Hierarchical Representations -VERSE: Versatile Graph Embeddings from Similarity Measures - Network Embedding as Matrix Factorization: Unifying DeepWalk, LINE, PTE, and node2vec - Adversarial Network Embedding - GraphGAN: Graph Representation Learning with Generative Adversarial Nets - Deep Graph Infomax Tech stack: Python, PyTorch, and various graph libraries. Why it matters: This repository is essential for researchers and practitioners interested in graph deep learning, providing a wealth of cutting-edge research in the field.

Overview: TradeMaster is an open-source platform designed for quantitative trading, powered by reinforcement learning, covering the full pipeline from design to deployment. Key features: - Reinforcement Learning (RL) enabled quantitative trading - Full pipeline support: design, implementation, evaluation, and deployment - Open-source and community-driven Tech stack: Python 3.9, Jupyter Notebook Why it matters: TradeMaster provides a comprehensive solution for quantitative traders, offering robust tools and a user-friendly interface for developing, testing, and deploying reinforcement learning algorithms in the trading domain.

### Overview: DeepCamera is an open-source AI camera skills platform that leverages local models like Qwen, DeepSeek, SmolVLM, and LLaVA for scene analysis, object detection, and person re-identification. The platform is designed for privacy and security, allowing cameras to run inference locally without exposing sensitive data. ### Key Features: - Local VLM video analysis using models like Qwen, DeepSeek, SmolVLM, and LLaVA. - Pluggable AI skills for various security and surveillance tasks. - Supports OpenAI, Google, Anthropic, or local AI models. - Runs on Mac Mini and AI PC. - Integrates with Slack, Discord, and Telegram for agent chat and notifications. ### Tech Stack: JavaScript ### Why It Matters: DeepCamera provides a robust and secure platform for AI-driven camera skills, making it easier to integrate AI into traditional CCTV systems. The platform's local model execution ensures maximum privacy and security, making it an attractive solution for businesses and individuals looking to enhance their security systems with AI capabilities.

### Overview: WindsurfAPI 是一个自托管的协议兼容网关，用于将 100+ Windsurf / Devin AI 模型（如 Claude、GPT、Gemini 等）转换为 OpenAI、Anthropic 和 Gemini 兼容的 API。 ### Key features: - **OpenAI API**: `POST /v1/chat/completions` - **OpenAI Completions API**: `POST /v1/completions` - **OpenAI Responses API**: `POST /v1/responses` - **Anthropic API**: `POST /v1/messages` - **Gemini API**: `POST /v1beta/models/*` ### Tech stack: Node.js ### Why it matters: 将 Windsurf / Devin 的 AI 模型转换为 OpenAI、Anthropic 和 Gemini 兼容的 API，实现零依赖的自托管解决方案，适用于企业级应用和自托管服务器环境。

### Overview: Neural Style is a TensorFlow implementation that uses convolutional neural networks to synthesize a pastiche by combining the content of one image with the style of another. This technique allows for the transfer of artistic styles from various artworks to a single content image, producing visually convincing results. ### Key Features: - **Neural Style Algorithm**: Separates and combines the content of one image with the style of another using convolutional neural networks (CNNs). - **Content / Style Tradeoff**: Allows for the control of the relative weight of style and content. - **Multiple Style Images**: Enables blending of multiple artistic styles to create a unique and diverse aesthetic. ### Tech Stack: Python ### Why It Matters: Neural Style is a powerful tool for creating visually appealing and stylistically diverse images. It has applications in various fields such as art restoration, design, and even in creating unique digital art pieces. The ability to transfer artistic styles from different sources to a single image opens up new possibilities in image manipulation and creative design.

Overview: Spice is a real-time analytics node that accelerates data-driven AI applications by integrating SQL, search, and LLM inference into your operational databases. It provides a portable, Rust-based engine that can connect to various data sources and serve data through standard APIs. Key features: - Connects to 30+ data sources - Accelerates datasets in memory or on local disk - Supports SQL queries, vector search, and LLM inference - Queries on a continuously refreshed dataset return in milliseconds - Queries on other datasets are served at the source Tech stack: Rust, Apache DataFusion, Apache Arrow, Vortex Why it matters: Spice simplifies the integration of data and AI into applications, enabling real-time analytics without the performance overhead of traditional data processing pipelines. Its ability to serve data through standard APIs makes it highly flexible and scalable, making it ideal for modern data-driven applications.

### Overview: DSH 桌面宠物是一款透明动画小桌宠，支持多开、大小位置随心配置，并内置 DIY 素材链，通过 AI 视频自造专属宠物。它不仅美观，还能实时响应 DSH 会话事件，提供丰富的动画效果，如呼吸、动作、切换状态等。这款插件是完整三件套项目，包含提示词、素材生成链和插件成品，可以自定义角色、动作和风格，从零生成自己的桌面宠物。 ### Key features: - **透明动画小桌宠**：支持多开、大小位置随心配置，内置 DIY 素材链，通过 AI 视频自造专属宠物。 - **实时响应会话事件**：宠物会根据 DSH 会话事件自动切换工作状态动画、播放余额动画等。 - **自定义功能**：支持更换角色、动作和风格，全流程可复现。 - **桌面应用兼容性**：适用于桌面应用（Electron 版）和 `dsh web`。 ### Tech stack: TypeScript ### Why it matters: DSH 桌面宠物是一款创新的透明动画小桌宠，不仅美观，还能实时响应 DSH 会话事件，提供丰富的动画效果。它通过 AI 视频自造，为用户提供个性化的宠物体验，是桌面应用和 `dsh web` 用户的必备工具。

### Overview: QuantMind is an open-source framework designed for quantitative finance, focusing on refining raw financial information into structured knowledge that can be easily retrieved and reasoned about. ### Key Features: - **Knowledge Engineering**: Converts any source into typed knowledge, ensuring provenance and persistence. - **Agent-Oriented Design**: Employs coding agents with human oversight, allowing for flexible and importable Python libraries. - **Replayable Preprocess**: Precise and deterministic preprocessing of sources, ensuring replayability and exact provenance. - **Typed Knowledge Structures**: Documents and news are structured as `Paper` and `News` objects, maintaining self-contained and time-queries-able artifacts. - **Retrieval Tools**: Supports various retrieval methods including RAG, library-based, and mind-based retrieval for deep research and data-MCP serving. ### Tech Stack: Python ### Why it Matters: QuantMind transforms raw financial information into structured knowledge, making it easier to extract actionable intelligence from quantitative finance data.

### Overview: EnvoyMesh is a decentralized, peer-to-peer mesh designed for autonomous AI agents. It offers a self-sovereign identity, secure communication, and AI-driven task delegation without relying on a central server. ### Key Features: - **Self-Sovereign Identity**: Ed25519 keys and cryptographic identity management. - **Mesh Protocol**: Peer-to-peer communication for discovery, chat, and task execution. - **Thin-Client Protocol**: Simplifies communication between home computer and phone. - **AI Agent**: Runs on hardware, following user policies. - **Security by Design**: Signed messages, policy-based trust, and end-to-end auditability. ### Tech Stack: TypeScript ### Why It Matters: EnvoyMesh provides a secure, decentralized, and self-sovereign way for AI agents to communicate and collaborate without relying on a central authority, ensuring privacy and control over personal data and AI interactions.

### Overview: Cheshire Cat AI is an AI agent microservice designed for educational and research purposes. It is built bottom-up to be easy to understand and extend using vibe coding and agentic engineering. ### Key Features: - **Educational and Research Use Case:** Cheshire Cat AI is primarily used for educational and research purposes, making it suitable for various institutions and companies. - **Easy to Understand and Extend:** The microservice is designed to be easy to understand and extend, allowing developers to quickly integrate and customize it. - **REST API Playground:** The project includes a REST API playground for developers to experiment and test functionalities. ### Tech Stack: Python ### Why it Matters: Cheshire Cat AI is a valuable resource for anyone looking to understand how AI agents work, making it an essential tool in the field of AI education and research. Repository: cheshire-cat-ai/core GitHub Stars: 3096 GitHub Description: AI agent microservice Roadmap & Contributing: Roadmaps are for amateurs. Proposals should be opened as issues, and only PRs should be made if assigned to you. The contributor guide is available in the repository. License: GPL3 Trademark: The Cheshire Cat AI logo and name are property of Piero Savastano.

Overview: MatMul-Free LM is a language model architecture that eliminates the need for Matrix Multiplication (MatMul) operations, providing a more efficient alternative for large-scale models. Key features: - Eliminates MatMul operations - Compatible with the 🤗 Transformers library - Available pre-trained models for 370M, 1.3B, and 2.7B parameter sizes Tech stack: Python Why it matters: This architecture is designed to be more efficient and scalable, potentially reducing computational costs and improving performance in large language models.

### Overview: Tract is a lightweight, no-nonsense, self-contained inference engine designed for efficient execution of ONNX and NNEF models across various platforms, including embedded ARM CPUs, NVIDIA GPUs, Apple GPUs, and web browsers via WebAssembly. It optimizes models for performance and is used in production at Sonos for wake-word and streaming speech-recognition tasks, as well as for large language models, text-to-image generation, and classical computer vision applications. ### Key Features: - **Efficient Inference**: Tract optimizes models for various hardware targets, ensuring optimal performance. - **Cross-Platform Compatibility**: Supports both CPU and GPU execution, as well as WebAssembly for browser usage. - **Model Optimization**: Automatically optimizes models for different hardware configurations to maximize performance. - **Cross-Platform Support**: Works seamlessly across different operating systems (Linux, macOS, Windows) and embedded systems. - **Lightweight**: Minimal footprint, making it suitable for resource-constrained environments. ### Tech Stack: - **Rust**: The core technology stack, providing the authoritative public API and internal crates. - **Rustc**: Rust version 1.91.0 or higher for Rust compatibility. - **MIT/Apache 2 License**: Open-source licensing. ### Why It Matters: Tract simplifies the process of deploying neural networks by abstracting away the complexities of hardware-specific optimizations and providing a unified API for inference across different platforms. This makes it an essential tool for developers looking to integrate AI models into their projects without worrying about the underlying hardware specifics. ### Repository

### Overview: Whishper is an open-source, 100% local audio transcription and subtitling suite with a full-featured web UI. Powered by whisper models! ### Key features: - 🗣️ **Transcribe any media** to text: audio, video, etc. - 📥 **Download transcriptions in many formats**: TXT, JSON, VTT, SRT or copy the raw text to your clipboard. - 🌐 **Translate your transcriptions** to any language supported by [Libretranslate](https://libretranslate.com). - ✍️ **Powerful subtitle editor** so you don't need to leave the UI! - 🏠 **100% Local**: transcription, translation and subtitle edition happen 100% on your machine (can even work offline!). - 🚀 **Fast**: uses FasterWhisper as the Whisper backend: get much faster transcription times on CPU! - 👍 **Quick and easy setup**: use the quick start script, or run through a few steps! - 🔥 **GPU support**: use your NVIDIA GPU to get even faster transcription times! - 🐎 **CPU support**: no GPU? No problem! Whishper can run on CPU too. ### Tech stack: Svelte ### Why it matters: Whishper is an essential tool for anyone looking to transcribe and subtitle audio and video content locally, ensuring privacy and control over their media. The open-source nature of Whishper means it can be used in a wide range of applications, from

Overview: CodeT5 and CodeT5+ are cutting-edge models designed by Salesforce Research for code understanding and generation. These models are part of a larger effort to enhance software development productivity using AI. Key features: - CodeT5: A unified pre-trained encoder-decoder model for code understanding and generation. - CodeT5+: An improved version of CodeT5 for better performance and flexibility. Tech stack: Python, utilizing Hugging Face's Transformers library. Why it matters: These models can significantly boost the productivity of software developers by automating tasks such as code generation, autocompletion, and summarization, thereby reducing the time and effort required for coding tasks. Repository: salesforce/CodeT5 GitHub stars: 3092 GitHub description: Home of CodeT5: Open Code LLMs for Code Understanding and Generation 

Overview: xLearn is a high-performance, easy-to-use, and scalable machine learning package that includes linear models, factorization machines, and field-aware factorization machines. It is designed for solving large-scale machine learning problems, particularly on sparse data. Key features: - Linear Model (LR) - Factorization Machines (FM) - Field-Aware Factorization Machines (FFM) - Python and CLI interface Tech stack: C++ Why it matters: xLearn provides a faster, more efficient, and easier-to-use alternative to existing machine learning libraries, making it suitable for large-scale sparse data problems. It supports out-of-core training, enabling the processing of very large datasets on a single machine. Repository: aksnzhy/xlearn GitHub stars: 3091

Overview: `prose` is a Go library designed for text processing, offering tokenization, part-of-speech tagging, named-entity extraction, readability metrics, and case conversion. Key features: - Tokenization - Segmentation - Tagging - Entities - Readability metrics Tech stack: Pure Go, no cgo, no external services. Why it matters: For developers working with Go, `prose` simplifies natural language processing tasks by providing a comprehensive library for text processing, making it easier to integrate NLP capabilities into Go applications.

### Overview: RamaLama simplifies the local serving and inference of AI models using familiar container-centric development patterns. ### Key Features: - **Eliminates Host System Configuration**: RamaLama automatically detects and pulls an accelerated container image specific to the host's GPUs, handling dependencies and hardware optimization. - **Supports Multiple AI Model Registries**: RamaLama integrates with OCI Container Registries, allowing easy access to various AI model registries. - **Containerized Models**: Models are treated similarly to container images, enabling common container commands for working with AI models. - **Secure Rootless Containers**: Runs models in rootless containers, isolating them from the host system, and securing data with no network access and temporary data removal. - **REST API and Chatbot Interface**: Supports both REST API and chatbot interfaces for interaction with AI models. ### Tech Stack: Python ### Why It Matters: RamaLama simplifies the local use and serving of AI models, making it easier for developers to work with AI in a familiar and secure environment.

### Overview: This repository is a curated list of robotics libraries and software, focusing on various aspects of robotics development. It includes middleware and frameworks, robot hardware and drivers, robot description and models, kinematics and dynamics, physics engines, simulators, motion planning and navigation, control, optimization, math and geometry, SLAM and state estimation, perception, grasping and manipulation, robot learning, and data visualization. ### Key Features: - 🟢 [ROS 2](https://github.com/ros2/ros2) - Version 2.0 of the Robot Operating System (ROS) software stack. - 🟢 [Dimensional](https://dimensional.org/) - Agentic operating system for physical space. - 🟢 [YARP](https://www.yarp.it/) - Communication and device interfaces applicable from humanoids to embedded devices. - 🟢 [ros2_medkit](https://selfpatch.github.io/ros2_medkit/) - Fault management and diagnostics gateway for ROS 2. - 🟢 [Robot Control Stack (RCS)](https://robotcontrolstack.org) - ROS-free Gymnasium framework for robot learning. - 🟡 [Micro-ROS for Arduino](https://github.com/kaiaai/micro_ros_arduino_kaiaai) - a Micro-ROS fork available in the Arduino Library Manager. - 🟢 [Rock](https://www.rock-robotics.org/) - Software framework for robotic systems. - 🟢 [ROS](https://www.ros.org/) - Flexible framework for