Projects

### Overview: Porda AI is the World's First Realtime onDevice AI Project designed to maintain modesty and religious values. It offers a powerful, innovative application for the Muslim community, allowing users to blur and cover non-mahram and haram objects directly on their screen in real-time, whether offline, online, or in any browser or video player. The app is built with Python and runs entirely offline, ensuring user privacy and data security. ### Key Features: - **Real-time Blurring and Covering:** Porda AI blurs and covers non-mahram and haram objects on the screen in real-time, ensuring modesty and religious adherence. - **Customizable Model:** The app features a custom AI model that optimally covers haram objects while preserving surrounding halal content, enhancing user experience. - **Secure and Privacy-Protected:** The application runs entirely offline, ensuring no data is sent or stored externally, and it guarantees user privacy and data security. - **User-Friendly Interface:** The app has an intuitive design and system tray integration for easy navigation and settings adjustment. - **Instant On/Off:** The app provides a shortcut key for instant on/off functionality. ### Tech Stack: Python ### Why it Matters: Porda AI empowers Muslim communities to adhere to their religious beliefs and maintain a more positive and sacred digital environment. It offers a powerful tool for users to protect themselves from inappropriate content and promotes a more respectful and modest digital presence. ### GitHub Stars: 323 ### Download and Installation: To download and install Porda AI, visit the

Overview: Lumen is a data visualization framework that turns natural language into SQL, charts, dashboards, and reports, making it easy to generate and share data pipelines and visualizations. Key features: - Generate SQL queries - Provide context and embeddings - Visualize data - Inspect, validate, and edit results - Summarize results and key insights - Extend Lumen with custom agents, tools, and analyses Tech stack: Python Why it matters: Lumen is a fully open-source and extensible agent-based framework that makes it easy to generate and share data pipelines and visualizations using natural language, making it ideal for non-programmers and those who want to perform complex analyses without needing to code.

Overview: This project implements Encoder-based Domain Tuning for Fast Personalization of Text-to-Image Models, focusing on the E4T-diffusion framework. It leverages Stable Diffusion and Stable UnCLIP for efficient text-to-image generation. Key features: - Pre-trained model for face images trained on CelebA-HQ+FFHQ - Domain-specific E4T pre-trained model for face images - Domain-tuning process for personalized image generation - Inference capabilities using placeholder tokens Tech stack: Python Why it matters: This project provides a practical implementation of E4T-diffusion, enabling faster and more personalized text-to-image generation, which is particularly useful for applications requiring high-quality image results with domain-specific styles. Repository: mkshing/e4t-diffusion GitHub stars: 323 GitHub description: Implementation of Encoder-based Domain Tuning for Fast Personalization of Text-to-Image Models using d🧨ffusers.

### Overview: This repository contains the companion code for the book "Machine Learning From Scratch," which offers a 5-stage framework for understanding and building core machine learning algorithms from scratch. The book compares these algorithms with industry-standard libraries like Scikit-learn and PyTorch, providing a comprehensive guide for anyone with basic Python and high-school math skills. The code is available in Jupyter Notebooks and can be run locally or imported into Google Colab. ### Key Features: - Companion code for "Machine Learning From Scratch" - 5-stage framework for understanding core ML algorithms - Compares algorithms with Scikit-learn and PyTorch - Uses NumPy for implementation - Includes notebooks for each chapter - Free access via GitHub ### Tech Stack: Jupyter Notebook ### Why it Matters: Anyone with basic Python and high-school math can fully understand and build core ML algorithms from scratch. The book provides a practical guide to the "black box" behind `fit()` and `predict()`, making it accessible to beginners and intermediate learners alike.

### Overview: This repository contains code for both my live course: "O'Reilly Live Online Training for AI Agents A-Z" and my video series: "Modern Automated AI Agents: Building Agentic AI to Perform Complex Tasks". The course provides a comprehensive guide to understanding, implementing, and managing AI agents, from foundational concepts to advanced topics like various frameworks and building agents from scratch using powerful prompt engineering techniques. It emphasizes practical application, guiding participants through hands-on exercises to implement and deploy AI agents, evaluate their performance, and iterate on their designs. The course covers key aspects such as cost projections, open versus closed source options, and best practices, equipping attendees with the knowledge to make informed decisions in their AI projects. ### Key Features: - Code for live course and video series - Comprehensive guide to AI agents - Foundational concepts to advanced topics - Frameworks like CrewAI, LangChain, and AutoGen - Building agents from scratch using prompt engineering techniques - Practical application with hands-on exercises - Informed decision-making with cost projections, open versus closed source options, and best practices ### Tech Stack: Jupyter Notebook ### Why it Matters: This course is essential for anyone looking to understand, implement, and manage AI agents in their projects. It provides a practical guide that equips participants with the knowledge to make informed decisions and effectively apply AI agents in real-world scenarios.

Overview: This repository provides a Jupyter Notebook Colab notebook for running the SANA-Sprint model, a variant of Stable Diffusion Hyper-SDXL, for generating high-resolution text-to-image prompts. Key features: - Colab notebook for Stable Diffusion Hyper-SDXL - Text-to-image generation using SANA-Sprint model - Multiple image results showcasing different prompts Tech stack: Jupyter Notebook Why it matters: This tool simplifies the process of generating high-resolution images from text descriptions, making it easier for developers and artists to experiment with and visualize creative prompts using advanced text-to-image models. GitHub stars: 323

### Overview: 本项目是机器学习相关代码和博客文章的集合，涵盖了从基础算法到竞赛代码的全面内容。主要涉及决策树、神经网络、感知机、支持向量机和逻辑斯谛回归模型的实现，以及文本分类和关键词提取的项目复现代码。 ### Key Features: - 决策树相关算法实现：包括ID3、C4.5、随机森林和Boosting算法的详细说明和实现。 - 神经网络实现：使用TensorFlow实现多层感知机和可视化训练过程中的数据记录。 - 感知机和支持向量机：详细解释两种具有相似性的机器学习算法。 - 逻辑斯谛回归模型和最大熵模型：代码实现和原理介绍。 - 文本分类：使用TextCNN模型实现句子分类。 - 文本相似度：项目复现蚂蚁金融NLP竞赛的文本语义相似度赛题。 - 关键词提取：实现中文文本关键词提取项目。 ### Tech Stack: Python ### Why it Matters: 这些项目和代码是机器学习领域的重要资源，适用于机器学习初学者和进阶者，帮助理解和实践机器学习算法。通过这些项目，读者可以学习到如何使用TensorFlow、Python等工具进行机器学习模型的实现和数据处理。

Overview: ActionWeaver is an AI application framework designed to simplify function calling with Language Models (LLMs) by relying solely on OpenAI and Pydantic. It supports both OpenAI API and Azure OpenAI service, making it highly versatile and user-friendly. Key features: - Function Calling as First-Class Citizen: Centralizes function-calling within the framework. - Extensibility: Integrates any Python code into the agent's toolbox with a single line of code. - Function Orchestration: Enables complex orchestration of function calls, including intricate hierarchies or chains. - Telemetry and Observability: Integrates with platforms like LangSmith for traceable applications. Tech stack: Python Why it matters: ActionWeaver simplifies the process of creating, orchestrating, and handling function calls, making it an essential tool for AI engineers looking to streamline their workflow and improve the reliability and efficiency of their projects. Repository: TengHu/ActionWeaver GitHub stars: 322 GitHub description: Make function calling with LLM easier README excerpt: [...](README excerpt)

### Overview: One-Prompt-One-Story: Free-Lunch Consistent Text-to-Image Generation Using a Single Prompt is a project that leverages a single prompt to generate consistent images across multiple subjects. This approach aims to provide a free lunch by offering a consistent image generation process using a single prompt, enhancing the efficiency and effectiveness of text-to-image generation. ### Key Features: - **Consistent Image Generation Code**: main.py - **Gradio Code**: app.py - **Benchmark Generation Code**: resource/gen_benchmark.py - **Consistency+ Benchmark**: [link](./resource/consistency+.yaml) - **Chinese Version Paper**: 🚩[Chinese Version Paper (new)](https://www.overleaf.com/read/pxyhqpsccfjk#20922a) - **Jittor Version**: 🚩[Jittor Version (new)](https://github.com/byliutao/1Prompt1Story/tree/main/jittor_version) ### Tech Stack: - Python ### Why It Matters: This project offers a significant improvement in text-to-image generation by providing a consistent image across multiple subjects using a single prompt, making it more efficient and effective.

Overview: The {chatgpt} R package provides a set of features to assist in R coding, including asking ChatGPT, commenting, completing code, creating unit tests, variable names, documentation, explaining code, finding issues, optimizing code, and refactoring code. Key features: - Ask ChatGPT - Comment selected code - Complete selected code - Create unit tests - Create variable name - Document code (in roxygen2 format) - Explain selected code - Find issues in the selected code - Optimize selected code - Refactor selected code Tech stack: R Why it matters: The {chatgpt} package streamlines the coding process by providing a set of tools to enhance productivity and accuracy in R programming, making it easier to write, test, and optimize code effectively. Repository: jcrodriguez1989/chatgpt GitHub stars: 322

Overview: 本项目是一个基于Node.js环境的中文分词包，旨在提供一个在任何JavaScript环境中可用的中文分词解决方案。它基于开源项目`node-segment`进行改造，增加了对Electron和浏览器的支持，并优化了多线程运行环境。 Key features: - 支持Node.js环境 - 支持Electron环境 - 支持浏览器环境 - 增加了对多线程运行环境的支持 - 改造了原项目代码，使其更易于使用和维护 - 提供了ES2015模块化版本 - 支持字典文件的树状优化（ESM模块） Tech stack: JavaScript Why it matters: 由于原项目`node-segment`在浏览器和Electron环境中运行时表现不佳，因此需要对代码进行重构以使其在这些环境中也能正常工作。此外，通过使用ES2015模块化和树状优化，可以提高代码的可读性和可维护性。这对于希望在任何JavaScript环境中使用中文分词的开发者来说是非常重要的。 Repository: linonetwo/segmentit

### Overview: PVDM is a PyTorch implementation of **Video Probabilistic Diffusion Models in Projected Latent Space**, presented at CVPR 2023. This model is designed for video generation tasks, specifically for generating videos from text inputs. It leverages diffusion models to achieve perceptual compression, making it suitable for applications where high-quality video synthesis is required. ### Key features: - **Video Generation**: Generates videos from textual prompts. - **Perceptual Compression**: Uses diffusion models to achieve high-quality video synthesis. - **Projected Latent Space**: Models are trained in a projected latent space for better performance. ### Tech stack: PyTorch ### Why it matters: This model is particularly useful in applications where high-quality video synthesis is required, such as in video editing, content creation, and other domains where realistic and high-resolution videos are needed. Its ability to generate videos from text inputs makes it versatile and applicable to a wide range of tasks.

### Overview: CenterSnap is a PyTorch implementation of a paper that presents a single-shot method for 3D shape reconstruction and categorical pose estimation. The repository includes the code for the ICRA'22 paper titled "Single-Shot Multi-Object 3D Shape Reconstruction and Categorical 6D Pose and Size Estimation". ### Key features: - Single-shot multi-object 3D shape reconstruction - Categorical 6D pose and size estimation ### Tech stack: PyTorch ### Why it matters: This repository provides a practical implementation of the proposed algorithm, allowing researchers and developers to easily reproduce the results and potentially adapt the method for their own applications. The code is open-source and can be used for academic research, as well as for commercial applications requiring real-time 3D reconstruction and pose estimation.

### Overview KodCode is a comprehensive framework for generating diverse, challenging, and verifiable synthetic datasets for coding tasks. It includes 12 distinct subsets covering various domains and difficulty levels, making it a valuable resource for both supervised fine-tuning and reinforcement learning tuning. ### Key Features - **Diverse Sources**: KodCode generates high-quality coding questions from multiple sources, including zero-shot generation, human-written assessment questions, code snippets, and technical documentation. - **Self-Verification**: It provides verifiable solutions and tests, supporting pytest and parallel execution. - **Style Converter**: Easy to convert between different styles of coding questions. ### Tech Stack Python ### Why It Matters KodCode is a significant contribution to the field of synthetic datasets for coding tasks, offering a comprehensive and diverse dataset that can be used for various applications, including fine-tuning models and reinforcement learning. Its ability to generate verifiable solutions and tests makes it particularly valuable in the development of AI-driven coding tools and systems.

### Overview: UGround is a comprehensive project that aims to enhance the ability of GUI agents to understand and interact with digital interfaces as humans do. This work, presented at ICLR'25, introduces UGround, a universal visual grounding model designed for GUI agents. ### Key Features: - **Universal Visual Grounding Model (UGround)**: UGround is a model that can be applied to a wide range of GUI tasks, including visual grounding, interaction, and understanding. - **Training Data**: The project includes comprehensive training data for the UGround model, including both text and image-based data. - **Model Weights**: Available weights for Qwen2-VL-based UGround-V1, including 2B, 7B, and 72B versions. - **Evaluation Suite**: A set of evaluation metrics and resources to help researchers test GUI agents and grounding models. - **Benchmark Performance**: Achieves state-of-the-art results on ScreenSpot-Pro benchmark, significantly improving upon previous models. ### Tech Stack: Python ### Why it Matters: UGround provides a powerful tool for GUI agents to interact with digital interfaces more naturally, making it invaluable for applications in the field of human-computer interaction. ### Repository: OSU-NLP-Group/UGround - **GitHub Stars**: 321 - **Primary Language**: Python - **GitHub Description**: [ICLR'25 Oral] UGround: Universal GUI Visual Grounding for GUI Agents

### Overview: Leanctx is a drop-in prompt compression solution for production LLM applications, offering a 10-40% reduction in token cost without compromising accuracy. It compresses requests uniformly across input, output, and context, reducing token usage by 18.7% on average, with 36.7% reduction on prose-heavy traffic. ### Key Features: - **Prompt Compression**: Automatically compresses input, output, and context to reduce token usage by up to 40%. - **Loss-Tolerance Routing**: Classifies prompts by their loss tolerance and compresses them differently based on the segment’s distortion. - **LLMLingua-2 Integration**: Routes prose-heavy prompts through LLMLingua-2 for additional compression. - **Open-Source and Local**: Runs on Open-source models, runs locally, and is MIT-licensed. ### Tech Stack: Python ### Why It Matters: Leanctx significantly reduces token costs without compromising accuracy, making it an essential tool for developers building production LLM applications.

Overview: Lightfeed Extractor is a TypeScript library designed to extract structured data from HTML, markdown, or plain text using language models (LLMs). It aims to provide robust, accurate results with minimal token usage, making it suitable for production data pipelines. Key features: - 🧹 LLM-ready Markdown conversion: Converts HTML to LLM-ready markdown with options to extract only main content and clean URLs. - 🚀 LLM Extraction function: Uses LLMs in JSON mode to extract structured data according to input Zod schema. Token usage limit and tracking included. - 🔗 JSON Recovery: Sanitizes and recovers failed JSON output, enhancing robustness, especially for complex schema extraction. - 🔝 URL Validation: Handles relative URLs, removes invalid ones, and repairs markdown-escaped links. - 🧹 Works with Playwright: Loads pages using Playwright to extract structured data from the HTML content. - 🔝 AI Browser Navigation: Navigates pages using natural language commands before extracting structured data. Tech stack: TypeScript Why it matters: Lightfeed Extractor is essential for developers looking to automate the extraction of structured data from web pages, making it easier to integrate into existing applications and pipelines. The library's ability to handle complex schemas and provide accurate results with minimal token usage makes it a valuable tool for data extraction tasks. Repository: lightfeed/extractor GitHub stars: 321 GitHub description: Use LLMs to robustly extract web data GitHub URL: https://github.com/lightfeed/extractor

### Overview This repository is a systematic survey of deep research (DR), organized across key components, training paradigms, and relevant benchmarks. It aims to provide a comprehensive overview of deep research systems, including foundational components, practical implementation techniques, important challenges, and future directions. ### Key Features - **Components**: Organized across four key components in DR, widely-used training paradigms, and relevant benchmark & resource. - **Tech Stack**: Unknown - **Why it Matters**: As the field of deep research continues to evolve rapidly, this survey is committed to continuously updating to reflect the latest progress in this area. ### Tech Stack - Unknown ### Why it Matters As the field of deep research continues to evolve rapidly, this survey is committed to continuously updating to reflect the latest progress in this area. ### Repository - **Repository**: mangopy/Deep-Research-Survey - **GitHub Stars**: 321 - **GitHub Description**: A Systematic Survey of Deep Research ### Latest News - **2025.11.25**: We release our survey [Deep Research: A systematic Survey](https://github.com/mangopy/Deep-Research-Survey/blob/main/Deep-Research-Survey.pdf). Thanks to my awesome co-authors. ### Table of Content - **Overview** - **Latest News** - **Reading List** - **Query Planning** - **Information Acquisition** - **Knowledge Boundary** - **Retrieval Timing** - **Agentic End-to-End Reinforcement

### Overview: This project involves the implementation of two deep learning algorithms for satellite image classification: U-Net and Pyramid Scene Parsing Network (PSPNet). The repository contains the complete code for training and testing these models using the IIT Bombay dataset. ### Key Features: - **U-Net Model**: Implements the U-Net architecture for satellite image classification. - **PSPNet Model**: Utilizes the PSPNet architecture for the same task. - **Training Data**: Includes training, validation, and testing data for the IIT Bombay dataset. - **Code Implementation**: Contains Python scripts for training, testing, and evaluating the models. - **Documentation**: Provides detailed documentation and images for better understanding. ### Tech Stack: - Python ### Why It Matters: This project showcases the application of deep learning techniques in satellite image classification, which is crucial for various applications such as environmental monitoring, urban planning, and disaster management. ### Repository: - **Repository**: `manideep2510/eye-in-the-sky` - **GitHub Stars**: 321 - **Primary Language**: Python - **GitHub Description**: Satellite Image Classification using semantic segmentation methods in deep learning - **Files**: - `main_unet.py`: Code for training the U-Net model. - `unet.py`: Implementation of U-Net layers. - `test_unet.py`: Code for testing, calculating accuracies, and saving predictions. - `Inter-IIT-CSRE`: Contains training, validation, and testing data. - `Comparison_Test.pdf

### Overview: This project provides a Python library and command-line tool for converting Common Crawl (CC) archives into a dataset of captions and documents. The library supports extracting text, audio, image, and video captions from CC archives, making it useful for building datasets in the field of machine learning. ### Key features: - Converts Common Crawl archives into a dataset of captions and documents. - Supports image, text, audio, and video captions. - Runs deduplication to save on output space and speed up the process. - Can be used as the first step in building a dataset, such as Laion5B. - Produces a large quantity of raw data that can be further filtered for better results. ### Tech stack: Python ### Why it matters: This tool simplifies the process of converting CC archives into a dataset, making it easier for researchers and data scientists to build high-quality datasets for machine learning applications. The ability to run in parallel and deduplicate reduces the time and resources required, making it a valuable contribution to the community. ### Repository: - Repository: rom1504/cc2dataset - GitHub stars: 321 - GitHub description: Easily convert common crawl to a dataset of caption and document. Image/text Audio/text Video/text, ... - Primary language: Python - GitHub description: Easily convert common crawl to a dataset of caption and document. Image/text Audio/text Video/text, ... ### Intended usage: - This tool produces a collection of link + caption. It is meant as the stage 1 of