Projects

### Overview: Pinvou Agent is an open-source desktop AI workspace designed for work, design, and coding. It integrates tools, files, knowledge, workflows, and real deliverables into a dedicated desktop environment, ensuring tasks end with a tangible result. ### Key Features: - **Multi-Session Workspace**: Persistent messages, tool calls, and artifacts from each session. - **Attachments**: Support for PDF, Office documents, images, and text. - **Artifact Panel**: Automatically collects and previews all created or edited files. - **Editable Markdown Artifacts**: Edit directly or select a passage for the agent to revise. - **Plan / YOLO Modes**: Review the plan first for complex work, or execute directly when the task is clear. - **Auxiliary Chat**: Ask side questions while a task runs, with a read-only view of the task's history. ### Tech Stack: Rust ### Why It Matters: Pinvou Agent revolutionizes how you work by integrating AI into your daily tasks, ensuring each project ends with a tangible result. Whether you're a designer, developer, or researcher, this workspace streamlines your workflow, making every task more efficient and productive.

### Overview: kordoc는 다양한 문서 형식을 Markdown로 변환하는 데 사용되는 CLI와 MCP 서버를 제공합니다. 이 도구는 HWP, HWPX, PDF, Office 문서를 구조화 데이터로 변환하여 라이브러리, CLI, MCP 서버로 사용할 수 있습니다. ### Key features: - **Markdown 변환**: 다양한 문서 형식을 Markdown으로 변환합니다. - **CLI 실행**: 설치 없이 `npx kordoc`으로 실행할 수 있습니다. - **MCP 서버**: 문서를 라이브러리, CLI, MCP 서버로 변환하여 사용할 수 있습니다. - **OCR 지원**: PDF 파일의 OCR 기능을 지원합니다. - **CLI 옵션**: 다양한 옵션을 제공하여 문서 변환을 쉽게 조절할 수 있습니다. - **CLI 전체**: CLI를 통해 모든 작업을 수행할 수 있습니다. ### Tech stack: - Node.js 20 이상 - TypeScript - npm v7.24.0 ### Why it matters: - **파일 변환**: 다양한 문서 형식을 Markdown으로 쉽게 변환할 수 있습니다. - **CLI 지원**: 설치 없이 CLI를 통해 다양한 작업을 수행할 수 있습니다. - **OCR**: PDF 파일의 OCR 기능을 지원하여 문서의 내용을 쉽게 변환할 수 있습니다. - **CLI 옵션**: 다양한 옵션을 제공하여 문서 변환을 쉽게 조절할 수 있습니다. -

### Overview: This project contains Jupyter notebooks for various algorithms discussed in Christopher Bishop's book "Pattern Recognition and Machine Learning" (PRML). It includes notebooks for chapters 1 through 14, covering topics such as Bayesian linear regression, support vector machines, Gaussian processes, and more. ### Key Features: - Jupyter notebooks for chapters 1-14 of PRML. - Includes notebooks for Bayesian linear regression, support vector machines, Gaussian processes, and more. - Replicas of graphs from the book for visual reference. ### Tech Stack: Python ### Why It Matters: This project serves as a practical implementation of the algorithms discussed in Christopher Bishop's book, providing a hands-on learning experience for developers interested in machine learning and pattern recognition.

Overview: This Python package provides fast and robust image stitching capabilities, built upon OpenCV's stitching module and inspired by the detailed stitching command-line tool. Key features: - Fast and robust image stitching - Supports multiple image formats - Easy installation via pip or Docker - CLI interface for command-line operations - Docker image for server environments - Python script for script-based usage - Stitching from multiple image filenames - Stitching from a single wildcard pattern - Stitching from preloaded images - Equivalent to the --affine CLI parameter - Equivalent to the -v/-verbose CLI parameter Tech stack: Python Why it matters: This package simplifies the process of stitching multiple images together, making it a valuable tool for various applications such as panoramic photography, video editing, and more. Repository: OpenStitching/stitching GitHub stars: 2636

### Overview: HarvestText 是一个专注于无监督或弱监督方法的文本挖掘和预处理工具，适用于小说分析、网络文本和专业文献等领域。它整合了领域知识，如类型和别名，以提高文本处理的效率和准确性。 ### Key features: - **文本清洗**: 处理URL、电子邮件、微博等特殊符号和格式，去除所有标点等。 - **实体链接**: 把别名、缩写与标准名联系起来。 - **命名实体识别**: 找到句子中的实体（如人名、地名、机构名）。 - **实体别名自动识别**: 从大量文本中自动识别实体及其可能的别名。 - **依存句法分析**: 分析语句中词语之间的主谓宾语等语法关系。 - **内置资源**: 通用停用词、通用情感词、领域词典（如IT、财经、饮食、法律）。 - **信息检索**: 统计特定实体出现的位置和次数。 - **新词发现**: 利用统计规律发现语料中可能被传统分词遗漏的特殊词汇。 - **字符拼音纠错**: 修正可能的错误拼写。 - **自动分段**: 对没有分段的文本自动分段，或者基于已有段落进一步组织/重新分段。 - **存取消除**: 本地保存模型再读取复用，或者消除当前模型的记录。 - **英语支持**: 本库主要旨在支持中文数据挖掘，但

### Overview: AudioLM is an implementation of the SOTA Language Modeling Approach to Audio Generation, developed by Google Research. It leverages Pytorch and extends the work for conditioning with classifier-free guidance, enabling text-to-audio or TTS functionalities. This repository includes a MIT-licensed version of SoundStream and is compatible with EnCodec, making it a versatile tool for audio generation and text-to-speech applications. ### Key Features: - **AudioLM Implementation**: A Pytorch-based implementation of the AudioLM model. - **Classifier-Free Guidance**: Conditioning the model with classifier-free guidance, allowing for text-to-audio or TTS functionalities. - **VALL-E Integration**: Training VALL-E from this repository, enabling the training of the VALL-E model. - **Compatibility**: Compatible with EnCodec and MIT-licensed SoundStream. ### Tech Stack: - Python ### Why It Matters: - Enables the training of VALL-E, a state-of-the-art text-to-audio model. - Facilitates the integration of SoundStream, a neural codec, for improved audio generation and TTS capabilities. ### GitHub Stars: 2625 ### Repository: - Repository: lucidrains/audiolm-pytorch - Primary Language: Python - GitHub Description: Implementation of AudioLM, a SOTA Language Modeling Approach to Audio Generation out of Google Research, in Pytorch. - GitHub Stars: 2625 ### Installation: ```bash $ pip install audiolm-pytorch ``` ### Usage: - **SoundStream &

Overview: VoiceMem is a project that introduces a new component to voice models: a soul, making them truly more understanding of you. It is built on a "streaming dual brain" architecture, providing precise, emotionally rich, and personable memory services with low latency and at an affordable price. The repository ensures that the project is permanently open-source. Key features: - Streamed dual brain architecture - Precise, emotionally rich, and personable memory services - Low latency through compression, layering, and streaming queries - Simple and practical, with single-token queries and fully decoupled architecture Tech stack: Python Why it matters: VoiceMem is designed to make voice models more intelligent and emotionally aware, making interactions more natural and engaging. The project's ability to handle low latency and provide affordable memory services makes it particularly valuable in applications where real-time responsiveness and personalization are crucial. GitHub stars: 2623 Repository: xzf-thu/VoiceMem

### Overview: This repository contains a curated list of deep learning papers focusing on industrial applications such as search, recommendation, and advertisement. The papers cover various techniques including embedding, matching, pre-ranking, ranking, post-ranking, relevance, LLM-based ranking, and reinforcement learning. The papers are categorized by their publication year and relevance to the industrial applications. ### Key Features: - **2013 (Google) (NIPS) [Word2vec] Distributed Representations of Words and Phrases and their Compositionality** - **2014 (KDD) [DeepWalk] DeepWalk - online learning of social representations** - **2015 (WWW) [LINE] LINE Large-scale Information Network Embedding** - **2016 (KDD) [Node2vec] node2vec - Scalable Feature Learning for Networks** - **2017 (ICLR) [GCN] Semi-supervised Classification with Graph Convolutional Networks** - **2017 (KDD) [Struc2vec] struc2vec - Learning Node Representations from Structural Identity** - **2017 (NIPS) [GraphSAGE] Inductive Representation Learning on Large Graphs** - **2018 (Alibaba) (KDD) *[Alibaba Embedding] Billion-scale Commodity Embedding for E-commerce Recommendation in Alibaba** - **2018 (ICLR) [GAT] Graph Attention Networks** ### Tech Stack: Python ### Why it Matters: These papers provide a comprehensive

### Overview: EdgeConnect is a generative image inpainting method that uses adversarial edge learning to fill in missing regions of an image. It excels at reproducing fine details by first generating edges of the missing regions and then using these edges as a priori to complete the image. ### Key features: - **Adversarial Edge Learning**: Combines an edge generator and an image completion network to hallucinate edges of the missing regions and use them as a priori for image inpainting. - **Two-Stage Approach**: EdgeConnect consists of two stages: an edge generator and an image completion network. - **Fine-Grained Details**: Produces better results in regions with fine details compared to traditional methods. ### Tech stack: Python ### Why it matters: EdgeConnect is significant because it addresses the limitations of traditional image inpainting methods by focusing on the edges of the missing regions, leading to more accurate and detailed reconstructions. This approach is particularly effective in applications where fine details are crucial, such as in artistic and historical restoration, where preserving the original texture and coloration is essential.

### Overview: Lobe Icons is a collection of AI/LLM brand logos designed for React and React Native apps. It provides 340+ AI model and provider logos in various styles, including mono, color, text, and avatar. The icons are static and can be used as SVG, PNG, or WEBP, making them compatible with any project. ### Key features: - **Every AI brand you need**: Models, providers, and AI apps in one place, updated regularly. - **Looks right everywhere**: Mono, color, text, and avatar styles that fit light and dark themes. - **Logo from an id**: Pass a model or provider ID like `gpt-4o` or `openai` and get the right logo. - **Small bundles**: Import only the brands you use. - **Use it anywhere**: React, React Native, or plain image URLs from a CDN. - **AI-native**: Your coding agent can learn it on its own from `skills.md` and `llms.txt`. ### Tech stack: TypeScript ### Why it matters: - **Every AI brand you need**: Models, providers, and AI apps in one place, updated regularly. - **Looks right everywhere**: Mono, color, text, and avatar styles that fit light and dark themes. - **Logo from an id**: Pass a model or provider ID like `gpt-4o` or `openai` and get the right logo. - **Small bundles**: Import only the brands you use. - **Use it anywhere**: React, React Native

### Overview This repository collects recent advances in the field of unifying large language models (LLMs) and knowledge graphs (KGs). It serves as a roadmap that summarizes three general frameworks: *1) KG-enhanced LLMs*, *2) LLMs-augmented KGs*, and *3) Synergized LLMs + KGs*. The repository aims to help researchers and practitioners gain a better understanding of this emerging field. ### Key Features - **KG-enhanced LLMs**: Techniques that integrate knowledge graph capabilities into LLMs to enhance reasoning. - **LLMs-augmented KGs**: Methods that leverage LLMs to augment knowledge graph capabilities. - **Synergized LLMs + KGs**: Frameworks that combine the strengths of both LLMs and KGs to achieve better reasoning. ### Tech Stack - Python - PyTorch - Graph Neural Networks (GNNs) ### Why It Matters The integration of LLMs and KGs can significantly enhance the ability of models to reason about complex, structured knowledge. This combination can lead to more accurate and contextually relevant predictions and insights, making it valuable in various applications such as language understanding, knowledge-based reasoning, and more. ### Repository Details - **Repository Name**: RManLuo/Awesome-LLM-KG - **GitHub Stars**: 2617 - **GitHub Description**: Awesome papers about unifying LLMs and KGs - **GitHub Description**: A collection of papers and resources about unifying large

### Overview: This repository contains a collection of papers, datasets, and tools related to crowd counting, a crucial aspect of computer vision and image processing. It serves as a comprehensive resource for researchers, developers, and enthusiasts interested in improving crowd counting techniques. ### Key Features: 1. **Datasets**: Includes datasets such as VSCrowd, FUDAN-UCC, RGBT-CC, JHU-CROWD++, and NWPU-Crowd, providing a wide range of data for benchmarking and research. 2. **Papers**: A curated list of recent advancements in crowd counting, including special issues, workshops, and challenges. 3. **Code**: Offers open-source code for crowd counting, including frameworks and models like C^3, CCLabeler, and YOLO-CROWD. 4. **Technical Blog**: Features detailed explanations of papers and methodologies, such as crowd counting from scratch and crowd counting with AI. 5. **GT Generation**: Provides tools and code for generating ground truth data, including density map generation from keypoints. ### Tech Stack: The repository utilizes PyTorch for its code, showcasing the versatility of machine learning frameworks in crowd counting applications. ### Why It Matters: Crowd counting is essential in various fields, including traffic management, public safety, and urban planning. This repository serves as a valuable resource for anyone looking to enhance their understanding and implementation of crowd counting techniques, making it a must-have for researchers and practitioners alike.

Overview: Meta-PyTorch TorchRec is a PyTorch library designed for building recommendation systems. Key features: - Efficient recommendation model training - Support for various recommendation algorithms - Integration with Meta-PyTorch for seamless model deployment Tech stack: Python Why it matters: Meta-PyTorch TorchRec simplifies the development of recommendation systems by providing a robust and efficient framework, making it a valuable tool for developers working in the field of recommendation systems.

Overview: This repository contains AI and Machine Learning contents designed for the Advanced Solutions Lab (ASL) bootcamps, focusing on various AI/ML topics such as model architectures, operationalizing models, and generative AI systems. Key features: - `asl_core`: Contains wide-ranging model architectures for multiple data modalities implemented in Tensorflow and Keras. - `asl_mlops`: Tools for operationalizing Tensorflow, Scikit-learn, and PyTorch models at scale using Google Cloud’s Agent Platform and Kubeflow pipelines. - `asl_genai`: Features generative AI and agent systems using frameworks like Gemini and Agentic. Tech stack: Python, Jupyter Notebook, Tensorflow, Keras, Scikit-learn, PyTorch, Google Cloud Platform (GCP), Google Cloud Agent Platform (GCPAP), Kubeflow, Gemini, Agentic. Why it matters: This repository provides comprehensive learning materials for AI/ML topics, enabling developers to operationalize models and create generative AI systems on Google Cloud, making it valuable for bootcamps and professional development.

### Overview: AgentField is an open-source control plane designed to build and run AI agents that can be called by any service in your stack, including frontends, backends, and other agents. AgentField simplifies the process of building and deploying AI logic by turning it into production infrastructure, including routing, coordination, memory management, and observability. It supports Python, Go, and TypeScript, and can scale from a single agent on a laptop to ten thousand agents in a single workflow. ### Key features: - **Build agents like APIs**: Write AI logic in Python, Go, or TypeScript. - **Run ten thousand of them like microservices**: Deploy agents as part of your existing infrastructure. - **Control plane handles fan-out, queues, and retries**: Ensures smooth operation and reliable execution. - **Production-ready**: Ready to deploy and run in a containerized environment. ### Tech stack: - Python, Go, TypeScript - Docker, Kubernetes, and other containerization tools ### Why it matters: - Simplifies the process of building and deploying AI logic. - Enables the creation of scalable and reliable AI services. - Supports integration with existing infrastructure and services. - Ensures smooth operation and reliable execution. ### Repository: - **Repository**: Agent-Field/agentfield - **GitHub stars**: 2,614 - **GitHub description**: Build, run, and scale AI agents like API and microservices ### Built with AgentField: - **Documentation**: Docs (https://agentfield.ai/docs/) - **Quick Start**: Quick Start (https://

### Overview: Apache Hamilton is a lightweight Python library designed for defining and executing data transformation pipelines. It allows data scientists and engineers to create modular, testable, and self-documenting dataflows that include lineage/tracing and metadata. ### Key Features: - **Expressive DAGs**: Define and modify data transformations using Python functions. - **Portable**: Runs on any Python environment, including scripts, notebooks, Airflow, and FastAPI. - **Modular**: ETL pipelines, ML workflows, LLM applications, RAG systems, BI dashboards, and the Apache Hamilton UI can all benefit from this library. - **Visualizable**: The DAGs can be easily visualized and understood. ### Tech Stack: Apache Hamilton is built on Python, leveraging its extensive ecosystem for data science and machine learning. ### Why It Matters: Apache Hamilton simplifies the process of defining and executing complex data transformations, making it easier to manage and maintain large-scale data pipelines. Its modular design ensures that each component of a dataflow is easy to modify and test, leading to more reliable and efficient applications. ### Repository: The repository for Apache Hamilton is hosted on GitHub, with 2,614 stars and a README that provides detailed instructions and examples for using the library. ### Primary Language: Apache Hamilton is primarily written in Jupyter Notebook, making it highly expressive and portable across various Python environments. ### GitHub Stars: The repository has received 2,614 stars, indicating a growing community interest and adoption of Apache Hamilton. ### Documentation: The library is well-documented, with a README that

Overview: SAG is a state-of-the-art retrieval architecture and knowledge base application built on SAG, designed to enhance the search and retrieval capabilities of documents and data. Key features: - Knowledge ingestion: Parses documents, splits them into chunks, embeds them, and extracts events and entities. - Search: Supports cross-source or source-scoped retrieval with Fast, Precise, and optional JEV reranking modes. - Source tracing: Allows users to trace results back to the original source chunks. - Knowledge graph: Enables users to inspect events, entities, and their associations. - Agent chat: Provides multi-turn answers grounded in selected sources, with clickable citations. - Integration: Supports self-hosted REST/OpenAPI, OpenAI-compatible chat, MCP, and the `zleap-sag` Python package. Tech stack: Python, SQLite, LanceDB, PostgreSQL/pgvector, OpenAPI, OpenAI-compatible chat, MCP, REST/OpenAPI, zleap-sag Python package. Why it matters: SAG provides a comprehensive knowledge base application that enhances the search and retrieval capabilities of documents and data, making it a valuable tool for individuals and agents alike.

### Overview: MTPLX is a native Mac app and command-line tool that runs local language models on Apple Silicon with the model's own multi-token prediction (MTP) heads. It supports Qwen 3.8 Flash Next, Qwen 3.8 27B, and Ternary Bonsai 2 27B, among others, achieving impressive speed and accuracy. ### Key Features: - **Model Support**: Runs Qwen 3.8 Flash Next, Qwen 3.8 27B, and Ternary Bonsai 2 27B. - **MTP Speculative Decoding**: Native MTP speculative decoding on Apple Silicon. - **Exact Sampling**: Exact sampling at the model's native temperature. - **OpenAI and Anthropic Compatibility**: Compatible with OpenAI and Anthropic local servers. - **Speed and Efficiency**: Measured at 125 tok/s in OpenCode on an M5 Max and 2.24x speed on an M5 Max, demonstrating significant performance improvements. ### Tech Stack: - Python ### Why It Matters: MTPLX offers the fastest way to run Qwen 3.8 on a Mac, achieving remarkable speed and efficiency. Its native MTP speculative decoding and exact sampling at the model's temperature make it a powerful tool for developers and researchers looking to optimize language model performance on Apple Silicon.

### Overview: Supabase Python Client is a comprehensive Python library for interacting with Supabase, a powerful database-as-a-service. It simplifies querying Postgres, handling user authentication, implementing security policies, creating edge functions, managing file storage, and enabling real-time data streaming. ### Key Features: - **Supabase Integration**: Directly integrates with Supabase for seamless database access. - **Postgres Querying**: Allows querying Postgres databases from Flask, Django, and FastAPI applications. - **User Authentication**: Provides robust user authentication and security policies. - **Edge Functions**: Enables the creation of edge functions for advanced data processing. - **File Storage**: Manages file storage for applications. - **Real-Time Data Streaming**: Supports real-time data streaming for applications. ### Tech Stack: Python ### Why it Matters: Supabase Python Client is a valuable tool for developers looking to integrate Supabase into their applications, offering a straightforward and efficient way to manage database interactions, user authentication, and real-time data handling. It is particularly useful for those new to Supabase or looking to streamline their development processes. Repository: supabase/supabase-py GitHub Stars: 2610

### Overview: This repository contains a collection of examples, end-to-end tutorials, and applications built using Liquid AI's foundational models (LFMs) and the LEAP SDK. It provides a range of tools and applications for developers to build with LFM and LEAP on laptops, mobile, and edge devices. ### Key Features: - **Desktop Apps**: Python and CLI applications for running LFM models. - **Browser Apps**: Zero-install applications running LFM models directly in the browser via WebGPU and ONNX Runtime Web. - **Mobile Apps**: Android and iOS applications using the LEAP SDK. - **Fine-Tuning Examples**: Detailed guides and examples for fine-tuning LFM models. - **Guides**: Comprehensive documentation and tutorials. - **Third-Party Apps Powered by LFM**: Applications built using LFM models. - **Community Projects**: Projects contributed by the community. - **Technical Deep Dives**: Detailed technical explanations and deep dives into the models. - **Contributing**: Instructions for contributing to the project. - **Support**: Resources for support and getting help. ### Tech Stack: Jupyter Notebook ### Why It Matters: This repository serves as a comprehensive resource for developers looking to integrate Liquid AI's models into their projects. It includes practical examples, tutorials, and applications, making it a valuable tool for anyone interested in leveraging the power of LFM and LEAP for AI development.