Projects

### Overview: Trafilatura is a comprehensive Python package designed to gather text from the Web and extract structured data, including text, metadata, and comments. It supports advanced web crawling, text discovery, and efficient processing of HTML files, making it a versatile tool for developers. ### Key Features: - **Advanced Web Crawling**: Supports sitemaps (TXT, XML) and feeds (ATOM, JSON, RSS) for comprehensive web discovery. - **Parallel Processing**: Efficiently processes live URLs and download queues, and supports previously downloaded HTML files and parsed HTML trees. - **Robust Extraction**: Extracts key elements like main text, metadata, formatting, and comments, with optional add-ons for language detection and speed optimizations. - **Output Formats**: Supports multiple formats including TXT, Markdown, CSV, JSON, HTML, and XML-TEI. ### Tech Stack: Python ### Why It Matters: Trafilatura is widely used across thousands of projects, including by HuggingFace, IBM, Microsoft Research, NVIDIA, the Allen Institute for AI, Stanford, and the Internet Archive, making it a reliable and efficient tool for developers looking to gather and extract meaningful data from the web.

Overview: This repository serves as a log of the skills and knowledge I am continuously acquiring to enhance my software engineering abilities and expand my knowledge across various technologies. Key features: - System Design: Covers foundational concepts in designing machine learning systems and data-intensive applications. - Maths: Includes courses on probability, statistics, hypothesis testing, and statistical thinking. - Data Structures and Algorithms: Focuses on algorithms and data structures, including linear algebra and eigenvectors. Tech stack: None specified. Why it matters: This log helps me stay updated with the latest technologies and methodologies in software engineering, ensuring I remain proficient and adaptable in the ever-evolving field of software development.

### Overview: This repository is a comprehensive reading list for research topics in multimodal machine learning, curated by Paul Liang. It includes tutorials, papers, and datasets from leading researchers in the field. ### Key features: - Tutorials on Multimodal Machine Learning - Course content from 11-777 Multimodal Machine Learning - Reading list for foundational and advanced topics in multimodal machine learning - Papers, slides, and videos from CVPR 2022 and NAACL 2022 - Comprehensive survey papers on principles, challenges, and open questions in multimodal machine learning ### Tech stack: Python, LaTeX, Markdown ### Why it matters: This repository serves as a valuable resource for researchers, students, and practitioners interested in the latest advancements and challenges in multimodal machine learning. It provides a curated collection of papers, tutorials, and datasets that cover a wide range of topics, from foundational principles to cutting-edge applications.

### Overview: TurboFieldfare is a custom Swift + Metal runtime designed to run the Gemma 4 26B-A4B model on any Apple Silicon Mac, including those with 8 GB of RAM. This model can be efficiently processed without loading the entire 14.3 GB model into memory, instead keeping the shared 1.35 GB core and FP16 KV cache in memory and streaming only the experts needed for each token from SSD. The runtime, streaming installer, CLI, and native Mac app are written in Swift and Metal, making it model-specific rather than a wrapper around MLX or llama.cpp. The curated experiment record summarizes 103 measured results across kernels, caching, I/O, prefill, and decode. ### Key features: - Model: Gemma 4 26B-A4B, 26B total parameters, about 3.88B active per token - Weights: MLX affine 4-bit, group 64; 8-bit router; 4-bit shared and routed experts - Memory: ~2 GB of weights and 4K KV cache - Storage: About 14.3 GB for the text model, plus about 1.1 GB for the optional image pack - Hardware: Apple Silicon Mac; 8 GB of RAM - Platform: macOS 26, Metal 4, Swift 6.2 - M2 measured decode: 5.1-6.3 tok/s on an 8 GB M2 MacBook Air

Overview: 飞桨产业级模型库是面向企业研发和快速集成的开源模型库，包含大量经过产业实践打磨的主流模型和在国际竞赛中夺冠的模型，提供端到端的开发套件，服务领域广泛，包括能源、金融、工业、农业等。 Key features: - 多个场景的开发套件，如语义理解、图像分类、目标检测、图像分割、文字识别、语音合成等。 - 飞桨官方模型超过600个，生态模型超过260个，持续更新中。 - 动态图和静态图模型，支持多种开发场景。 - 飞桨官方实现并提供持续技术支持及答疑。 - 与飞桨核心框架版本对齐，经过充分测试。 Tech stack: Python。 Why it matters: 飞桨产业级模型库是面向企业研发和快速集成的开源模型库，提供端到端的开发套件，满足企业低成本开发和快速集成的需求。 Repository: PaddlePaddle/models Primary language: Python GitHub stars: 6931 GitHub description: 官方模型库包含大量经过产业实践打磨的主流模型和在国际竞赛中夺冠的模型，提供面向语义理解、图像分类、目标检测、图像分割、文字识别、语音合成等场景的多个端到端开发套件，服务领域广泛，包括能源、金融、工业、农业等。

### Overview: Donut is an OCR-free document understanding transformer that achieves state-of-the-art performance across various visual document understanding tasks. It is based on the academic paper "OCR-free Document Understanding Transformer" from ECCV 2022, providing detailed experimental results and analyses. ### Key Features: - **OCR-free Document Understanding Transformer (Donut)**: A method that utilizes an OCR-free end-to-end Transformer model to perform document understanding tasks. - **Synthetic Document Generator (SynthDoG)**: A tool that helps the model pre-train to be flexible on various languages and domains. - **Official Implementation**: Provides full experimental results and analyses, including pre-trained models and web demos. ### Tech Stack: - Python ### Why It Matters: Donut and SynthDoG offer a new approach to document understanding that does not require off-the-shelf OCR engines, yet still achieves high performance. This method is particularly useful for applications where OCR is not feasible or desirable, such as in environments with limited access to OCR tools. ### Repository: - **Repository**: clovaai/donut - **Stars**: 6927 - **Description**: Official Implementation of OCR-free Document Understanding Transformer (Donut) and Synthetic Document Generator (SynthDoG), ECCV 2022 - **Primary Language**: Python - **Downloads**: 2.2k - **PyPI**: pip install donut-python - **Demo**: Gradio web demos available at [Gradio space web demo](https://huggingface.co/sp

Overview: This project introduces a PyTorch implementation of a method for high-resolution image-to-image translation, specifically focusing on synthesizing and manipulating 2048x1024 images using conditional Generative Adversarial Networks (GANs). Key features: - Synthesizing and manipulating 2048x1024 images with conditional GANs - Utilizing semantic label maps to generate photo-realistic images - Generating portraits from face label maps - Interactive editing capabilities Tech stack: The project is implemented using PyTorch. Why it matters: This method offers significant advancements in high-resolution image manipulation, making it particularly useful in applications such as creating photorealistic streetview images and generating realistic portraits from facial feature maps.

Overview: SQLBot 是一款基于大语言模型和 RAG 技术的智能问数系统，旨在提供对话式数据分析（ChatBI）功能，简化数据提取和可视化过程。 Key features: - 实现对话式数据分析（ChatBI） - 快速提炼获取所需的数据信息及可视化图表 - 支持进一步智能分析 - 开箱即用，无需复杂开发 - 提供安全可控的工作空间资源隔离机制 - 支持细粒度数据权限配置 - 支持多种集成方式，快速嵌入到 n8n、Dify、MaxKB、DataEase 等应用 - 高效运营，基于用户交互数据持续迭代优化 Tech stack: JavaScript Why it matters: SQLBot 提供了一种简单且高效的方式来处理复杂的数据分析任务，通过结合大语言模型和 RAG 技术，大大提高了数据提取和分析的效率，同时确保了数据的安全性和合规性。

### Overview: AppAgent is an LLM-based multimodal agent framework designed to operate smartphone apps. It simplifies the interaction process by mimicking human-like actions like tapping and swiping, allowing it to navigate and use new apps without requiring backend system access. ### Key Features: - **Simplified Action Space**: Enables the agent to interact with apps through a simplified interface, similar to how humans operate smartphones. - **Autonomous Exploration**: Learns to navigate and use new apps through autonomous exploration or by observing human demonstrations. - **Knowledge Base**: Generates a knowledge base that the agent uses to execute complex tasks across different applications. - **GUI Agent with Evolving Mechanism**: The next-generation GUI Agent with an evolving mechanism. ### Tech Stack: Python ### Why It Matters: AppAgent simplifies app interaction, broadening its applicability across diverse apps without the need for backend system access. This makes it more accessible and user-friendly, enhancing the overall user experience.

Overview: PostgresML is a Postgres extension that integrates machine learning and AI functionalities directly within the database, leveraging GPU acceleration for faster computations and model inference. Key features: - In-Database ML/AI - GPU Acceleration - Large Language Models - RAG Pipeline - Vector Search - Diverse ML Algorithms - High Performance - Scalability - NLP Tasks - Security - Seamless Integration Tech stack: Rust Why it matters: PostgresML enables efficient, reliable, and scalable ML/AI operations directly within the database, eliminating the need for separate systems and data transfers, making it ideal for applications that require both data storage and machine learning inference. Repository: postgresml/postgresml GitHub stars: 6823 GitHub description: Postgres with GPUs for ML/AI apps.

Overview: annyang is a tiny, lightweight JavaScript library for speech recognition that allows users to control websites with voice commands. It's free, open-source, and doesn't require any dependencies. Key features: - Speech recognition for site control - No dependencies, weighs only 2 KB - MIT license Tech stack: TypeScript Why it matters: annyang simplifies integrating speech recognition into web applications, making it accessible for developers looking to enhance user interaction through voice commands.

### Overview: 记录每天整理的计算机视觉/深度学习/机器学习相关方向的论文。 ### Key features: - 每日更新计算机视觉/深度学习/机器学习领域的优质论文、项目和应用速递。 - 包括2017-2023年的CVPR、ICLR、ICML等顶会顶刊论文。 - 涵盖目标检测、语义分割、目标跟踪、Transformer、多模态、大模型、NeRF、扩散模型、深度估计、超分辨率、3D目标检测、CNN、GAN、竞赛解决方案、人脸识别、数据增广、人脸检测、数据集、NAS、AutoML、图像分割、SLAM、实例分割、人体姿态估计、视频目标分割、Re-ID、医学图像分割、显著性目标检测、自动驾驶、人群密度估计、PyTorch、人脸、车道线检测、去雾、全景分割、行人检测、文本检测、OCR、6D姿态估计、边缘检测、场景文本检测、视频实例分割、3D点云、模型压缩、人脸对齐、超分辨、去噪、强化学习、行为识别、OpenCV、场景文本识别、去雨、机器学习、风格迁移、视频目标检测、去模糊、显著性检测、剪枝、活体检测、人脸关键点检测、3D目标跟踪、视频修复、人脸表情识别、时序动作检测、图像检索、异常检测等。 ### Tech stack: Python, Markdown, Git, GitHub ### Why it

### Overview: Qwen-VL is the official repository of Qwen-VL, a large language model developed by Alibaba Cloud. This model is designed to enhance language understanding and reasoning capabilities, particularly in visual tasks. It supports access through various channels, including the web, app, and APIs, making it accessible to developers and researchers alike. ### Key Features: - **Qwen-VL-Plus**: A significant upgrade with enhanced visual reasoning capabilities, detailed image recognition, and support for high-resolution images. - **Qwen-VL-Max**: Offers improved visual perception and cognitive understanding, particularly in complex tasks, surpassing previous models in multiple multimodal tasks. ### Tech Stack: Python ### Why It Matters: Qwen-VL is a powerful tool for developers looking to integrate advanced visual language models into their projects. Its enhanced capabilities make it suitable for applications requiring high-resolution image recognition and detailed text analysis, setting new benchmarks in the field of multimodal AI research and development.

### Overview: 本书深入探讨了人工智能基础设施（AI Infra）的设计与实现，从硬件约束出发，结合模型架构进行系统设计。通过详细的推理与训练负载分析、加速器架构优化、算子与运行时管理、数据中心网络设计、推理优化、分布式推理、训练系统安排以及资源调度与运行环境的考量，本书旨在帮助读者理解如何在有限的资源下高效地开发基于模型的应用。 ### Key features: - **硬件约束与模型架构结合**：从硬件性能出发，量化推导系统设计。 - **推理与训练负载分析**：详细解析任务阶段、状态寿命、计算与存储需求。 - **加速器架构优化**：在计算、存储、带宽、功耗与成本之间取舍。 - **算子与运行时管理**：融合、复用、并发与调度减少执行开销。 - **数据中心网络设计**：网络带宽、通信方式和拥塞影响计算效率。 - **推理优化**：批处理、KV管理、卸载与推测解码何时有效。 - **分布式推理**：计算与状态放置、扩缩容与恢复策略。 - **训练系统安排**：显存、通信和重算安排让训练更高效。 - **资源调度与运行环境**：模型服务与工具环境共享资源，减少等待时间。 ### Tech stack: Python 本书使用Python进行代码实现和数据分析。 ### Why it matters: 本书通过从硬件约束出发，结合模型架构进行系统设计，帮助读者

### Overview: FlashInfer is a high-performance library and kernel generator designed for inference, delivering state-of-the-art performance across diverse GPU architectures. ### Key Features: - **Attention Kernels**: Optimized for prefill, decode, and mixed batching scenarios. - **Multiple Backends**: Automatically selects the best backend for your hardware and workload. - **Modern Architecture Support**: Supports SM75 (Turing) and later GPUs. - **Low-Precision Compute**: Supports FP8 and FP4 quantization for attention, GEMM, and MoE operations. - **Production-Ready**: Compatible with CUDAGraph and torch.compile for low-latency serving. ### Tech Stack: Cuda ### Why It Matters: FlashInfer provides optimized kernels for inference, enabling high performance across various GPU architectures. It supports multiple backends and is compatible with modern hardware, making it a valuable tool for developers working on LLM serving.

### Overview: This project is a Model Context Protocol (MCP) server for WhatsApp, allowing users to access, search, and send messages directly from their personal WhatsApp accounts. It connects to the WhatsApp web multidevice API using the `whatsmeow` library and stores messages locally in a SQLite database, only sending them to an LLM when accessed through tools controlled by the user. ### Key Features: - **Personal Access**: Users can search and read their WhatsApp messages, including images, videos, documents, and audio. - **Contacts Search**: The ability to search and view contacts. - **Media Files**: Send and receive media files such as images, videos, documents, and audio messages. - **Local Storage**: Messages are stored locally in a SQLite database, ensuring privacy and security. ### Tech Stack: Go ### Why it Matters: This project is designed to be a secure and efficient way to access WhatsApp messages without the need for a web browser or third-party apps, making it ideal for users who prefer a more direct and secure experience. The use of a local SQLite database ensures that user data is stored and managed privately, enhancing privacy and security.

Overview: Mycroft Core is the core component of the Mycroft AI platform, serving as the foundation for all voice assistant functionalities. Key features: - Microphone and audio processing - Speech-to-Text (STT) and Text-to-Speech (TTS) capabilities - Skill management and execution - Configuration and management of Mycroft devices and accounts Tech stack: Python Why it matters: Mycroft Core is essential for the functionality and performance of Mycroft AI, providing the core processing and interaction capabilities that enable the voice assistant to understand and respond to user commands.

### Overview: Scikit-image is a powerful Python library for image processing, providing a comprehensive suite of tools for tasks such as filtering, segmentation, and feature extraction. It is widely used in scientific research, data analysis, and machine learning applications. ### Key Features: - **Image Processing Tools:** A wide range of image processing techniques including filtering, segmentation, and feature extraction. - **Integration with Scientific Python:** Built on top of the Scientific Python ecosystem, making it easy to integrate with other scientific libraries. - **Community and Documentation:** Active community support and detailed documentation, including forums and a user forum. ### Tech Stack: Python ### Why it Matters: Scikit-image is essential for researchers and developers working in fields such as computer vision, medical imaging, and data science. Its comprehensive set of tools and community support make it a go-to library for image processing tasks across various applications.

### Overview: NLP.js is an open-source library designed for building natural language processing (NLP) applications in Node.js. It offers a wide range of functionalities including language detection, sentiment analysis, entity recognition, and more, with support for multiple languages and plugins for enhanced functionality. ### Key Features: - **Language Detection**: Determines the language of a given phrase. - **Levenshtein Distance**: Calculates the minimum number of single-character edits required to change one phrase into another. - **Substring Matching**: Finds the best substring of a string with the lowest Levenshtein distance to a given pattern. - **Stemmers and Tokenizers**: Provides stemmers and tokenizers for various languages. - **Sentiment Analysis**: Analyzes phrases for sentiment, including negation support. - **Named Entity Recognition**: Recognizes named entities in text, supporting multi-language and similar string handling. - **Classifier**: Classifies utterances into predefined intents. - **Natural Language Generation**: Manages answers for a given utterance, including entity extraction, intent classification, and sentiment analysis. - **Support for 40 Native and 104 BERT-supported Languages**: Includes support for various languages, including fantasy languages. - **Plugin System**: Allows customization and extension of the library with plugins. - **Container System**: Manages plugins, settings, and pipelines for a more organized and modular approach. - **Connector System**: Provides connectors for integrating with other platforms like Microsoft Bot Framework and Direct Line. ### Tech Stack: Node.js ### Why It Matters: NLP.js simplifies the

Overview: OpenMVG is an open-source library for Multiple View Geometry, providing tools for 3D computer vision and Structure from Motion. It is designed to be easy to read, learn, modify, and use, with a strong focus on simplicity and maintainability. Key features: - Multiple View Geometry (MVG) library - Image manipulation and feature processing - Feature tracking and camera models - Multiple-view geometry algorithms - Structure-from-Motion (SfM) algorithms - Robust estimation techniques - State-of-the-art and classic algorithms Tech stack: C++ Why it matters: OpenMVG simplifies reproducible research in 3D reconstruction from images/photogrammetry, making it easier to build and trust larger systems. The library is designed for ease of use, with a strict test-driven development and samples that allow for accurate implementation of state-of-the-art and classic algorithms.