Projects

Overview: CommandCenter is a versatile StarCraft AI bot that supports both StarCraft: Broodwar and StarCraft II, utilizing the BWAPI and Blizzard's StarCraft II AI API. It provides a modular architecture for easy modification and customization. Key features: - Plays both Starcraft games with the same source code - Supports all three StarCraft races with generalized micro-controllers for combat units - Performs online map analysis, extracting information such as base locations and expansions - Tracks previously seen enemy units and their locations - Manages resource gathering and worker allocation/buying - Carries out predefined build-orders from a configuration file - Allows easy creation and modification of build-orders in-game - Scouts the map with a worker unit - Commences attacks when a specific condition is met (12 combat units by default) - Forms squads of units following specific orders - Supports 1v1 gameplay Tech stack: C++ Why it matters: CommandCenter is a valuable tool for StarCraft enthusiasts and AI developers, offering a robust and customizable framework for creating AI bots that can adapt to various strategies and scenarios.

Overview: FashionCLIP is a CLIP-like model fine-tuned specifically for the fashion domain, designed to enhance the representation and classification of fashion items. Key features: - Fine-tuned CLIP model for fashion classification - Capable of extracting features and classifying fashion items - Updated with the Laion CLIP-ViT-B-32-laion2B-s34B-b79K checkpoint for improved performance Tech stack: Python Why it matters: FashionCLIP offers a specialized solution for the fashion industry, improving representation and classification of fashion items through its fine-tuned CLIP model, making it particularly useful for applications requiring high accuracy in fashion-related tasks. Repository: patrickjohncyh/fashion-clip Primary language: Python GitHub stars: 543 GitHub description: FashionCLIP is a CLIP-like model fine-tuned for the fashion domain.

Overview: Magic-Insight is a one-stop data intelligence agent that provides insights from various data formats, including documents, databases, business systems, and images. Key features: - Supports multiple data formats - Provides comprehensive functionality for data generation, model fine-tuning, and deployment - Offers low usage threshold and high accuracy Tech stack: Python Why it matters: Magic-Insight simplifies data analysis by integrating AI capabilities into a user-friendly interface, making it accessible to developers and businesses alike. Repository: tianlan-ltd/Magic-Insight Primary language: Python GitHub stars: 546 GitHub description: One-stop data intelligence agent, providing insights from all mainstream data formats in a single dialogue box, including documents, databases, business systems, and images.一站式数据智能体，一个对话框提供所有主流格式数据的见解，包括文档、数据库、业务系统和图像。 README excerpt: # Introduction <a href="https://github.com/chaos-water/Magic-BI/blob/main/README_zh.md" style="color: blue; font-size: 18px;">中文请点这里</a><br/> # Introduction Magic-BI is an AI-based fully automated ChatBI product that currently supports SQL databases. Magic-BI is open-source and can be deployed in a fully private or semi-private manner, which maximizes user privacy protection while lowering the usage threshold. # Competitor Comparison | | Magic-BI | Other Chat BI Products | |:---------|:-------------------------------------------------------------------------------------------------------------|

### Overview: This repository is a curated list of resources for embedded AI, focusing on hardware, software, frameworks, and other tools for deploying AI at the edge. It includes resources for low-power, constrained devices such as STM32 microcontrollers, TensorFlow-enabled cameras, edge AI processors, and more. ### Key Features: - **Hardware**: Resources for low-power, constrained devices like STM32 microcontrollers and edge AI processors. - **Software**: TensorFlow-enabled cameras, edge AI processors, and low-power microcontrollers. - **Contributing**: Resources for contributing to the edge AI community. - **License**: The repository is licensed under the MIT License. ### Tech Stack: - TensorFlow - MicroPython - Edge TPU - NVIDIA Jetson - OpenMV - JeVois - Intel Movidius - NVIDIA Jetson - Artificial Intelligence Radio - Transceiver (AIR-T) - Kendryte K210 - Kendryte K510 - GreenWaves GAP8 - GreenWaves GAP9 - Ultra96 - Google Coral - Gyrfalcon Technology Lighspeeur - ARM microNPU - ESP32-S3 ### Why it Matters: This repository serves as a comprehensive resource for developers looking to integrate AI at the edge, focusing on hardware and software solutions for constrained environments. It provides a curated list of tools and resources that are essential for deploying AI applications in low-power, resource-constrained devices.

Overview: This repository is a curated list of popular deep learning models related to classification, segmentation, and detection problems. Key features: - Classification models - Segmentation models - Detection models Tech stack: Unknown Why it matters: This repository provides a comprehensive overview of popular deep learning models used in computer vision tasks, including classification, segmentation, and detection, with evaluation metrics from papers.

Overview: This repository is a curated list of papers on Language-Model-as-a-Service (LMaaS), focusing on papers that adapt large language models (LLMs) to serve various downstream tasks without requiring access to the model's parameters and gradients. Key features: - Text Prompt - In-Context Learning - Black-Box Optimization - Feature-based Learning - Data Generation Tech stack: Python, Markdown, and GitHub Pages. Why it matters: This repository serves as a valuable resource for researchers and practitioners interested in leveraging pre-trained LLMs for a wide range of applications, making them more accessible and deployment-efficient compared to fine-tuning.

### Overview: ComfyUI-H3-FaceRefine is a ComfyUI node designed to refine and improve the quality of small faces in MiniMax H3 video. It detects faces on every frame, crops them to fill a canvas, and then re-generates them using H3, finally stitching them back into the original video. ### Key features: - Face tracking and cropping - Improved face quality - Per-frame face tracking and re-generation with H3 - Smooth interpolation and composite fade for better temporal behavior ### Tech stack: Python ### Why it matters: This node addresses a common issue in MiniMax H3 video where small faces are poorly rendered due to the head occupying a small fraction of the frame. By refining these faces, it ensures better quality and continuity in the video, making it more suitable for various applications.

### Overview: Mengzi 是一个轻量级但强大的预训练语言模型系列，旨在通过融入语言学信息和加速训练方法，提供更优的部署和工业落地性能。 ### Key features: - **Mengzi-BERT-base**: 110M 参数量，适用于文本分类、实体识别、关系抽取、阅读理解等任务。 - **Mengzi-BERT-L6-H768**: 60M 参数量，由 Mengzi-BERT-large 蒸馏获得，适用于金融领域的自然语言理解类任务。 - **Mengzi-T5-base**: 220M 参数量，适用于可控文本生成任务，与 GPT 定位不同，适合文本续写。 - **Mengzi-T5-base-MT**: 220M 参数量，提供 Zero-Shot、Few-Shot 能力。 - **Mengzi-Oscar-base**: 110M 参数量，基于 Mengzi-BERT-base 的多模态模型，适用于图片描述、图文互检等任务。 - **Mengzi-GPT-neo-base**: 125M 参数量，基于中文语料从头训练，适合作为文本续写类任务的基准模型。 - **BLOOM-389m-zh**: 389M 参数量，基于中文语料对多语言版本进行裁剪的 BLOOM 模型，降低了对显存的需求。 - **BLOOM-800m-z

### Overview: MiniMax-M2.1 is a state-of-the-art model designed for real-world development and agent applications. It offers enhanced robustness in coding, tool use, instruction following, and long-horizon planning, making it transparent, controllable, and accessible. ### Key Features: - **Enhanced Robustness**: Optimized specifically for robustness in coding, tool use, instruction following, and long-horizon planning. - **Transparent and Controllable**: Empowers developers to build the next generation of autonomous applications while maintaining transparency and controllability. - **Open-Source Model Weights**: Available for local deployment and use, allowing for easy integration and customization. ### Tech Stack: - **Model**: MiniMax-M2.1 - **Language**: Unknown ### Why It Matters: MiniMax-M2.1 is a significant step forward in democratizing high-performance agent capabilities, making it accessible to developers worldwide. Its robustness and transparency ensure that AI is not just a tool but a partner in innovation. ### Repository: - **Repository**: MiniMax-AI/MiniMax-M2.1 - **GitHub Stars**: 543 - **GitHub Description**: MiniMax M2.1, a SOTA model for real-world dev & agents. ### Join Our Community: - **WeChat**: Join the community on our WeChat platform. - **Discord**: Connect with us on our Discord community. ### Benchmarks: - **SWE-bench Verified**: 74.0 vs 69.4 - **

### Overview: This repository is a curated collection of essential AI & Machine Learning books, designed for practitioners, researchers, and learners at all levels. It provides strong theoretical foundations, practical industry insights, and production-level engineering perspectives to support continuous learning across the AI stack. ### Key Features: - **Books**: Over 10 books covering various topics such as Machine Learning, Data Science, Deep Learning, and more. - **Theoretical Foundations**: Strong theoretical bases for AI and ML concepts. - **Practical Insights**: Industry-oriented insights and production-level engineering perspectives. - **Intermediate Level**: Suitable for intermediate learners and developers. - **Advanced Topics**: Covers advanced topics like Active Learning Workflows and Foundation-Model-Driven Systems. ### Tech Stack: The repository is written in Markdown and uses GitHub for version control and hosting. ### Why It Matters: This curated collection of books serves as a knowledge hub for AI enthusiasts, providing essential resources for practitioners, researchers, and learners across various levels of expertise. It supports continuous learning and development in the rapidly evolving field of AI and ML. ### Repository Overview: - **Stars**: 544 - **Description**: This repository is for all those AI enthusiasts who love to read books and learn. - **License**: The repository is licensed under the terms specified in the LICENSE file. ### Book References: - **Title**: Designing Machine Learning Systems - **Author(s)**: Chip Huyen - **Topic Area**: Machine Learning - **Key Focus**: Production ML systems & iteration - **Level**: Intermediate

### Overview: T2F: Text-to-Face Generation using Deep Learning ### Key features: - Combines StackGAN and ProGAN architectures for text-to-face generation. - Uses Face2Text dataset for training, containing 400 facial images and textual captions. - Generates realistic faces from textual descriptions using PyTorch framework. - Updates to use MSG-GAN for image generation module, replacing ProGAN. ### Tech stack: PyTorch ### Why it matters: T2F offers a unique approach to text-to-face generation, utilizing state-of-the-art deep learning techniques to create lifelike images from textual descriptions. This project is an excellent resource for developers looking to explore advanced text-to-image generation methods.

Overview: PyCNN is a Python library for implementing Cellular Neural Networks (CNN) for image processing applications. Key features: - CNN implementation for image processing - Supports Python 2.7 and Python 3.3 - Includes a template library for common image processing techniques - Used in research papers and a demo at the 14th Cellular Nanoscale Networks and Applications (CNNA) Conference 2014 Tech stack: Python Why it matters: PyCNN provides a practical solution for image processing tasks using CNNs, offering a parallel computing paradigm for real-time processing.

### Overview: AWS Generative AI CDK Constructs are sample implementations of AWS CDK for common generative AI patterns. These constructs aim to help developers quickly build generative AI solutions using pattern-based definitions for their architecture. ### Key features: - **Pattern-based Definitions**: AWS Generative AI CDK Constructs provide high-level, multi-service abstractions of AWS CDK constructs. - **Default Configurations**: Patterns are based on well-architected best practices, ensuring predictable and repeatable infrastructure. - **Logical Modules**: Each pattern is organized into logical modules using object-oriented techniques. ### Tech stack: TypeScript ### Why it matters: These constructs are designed to help developers build generative AI solutions efficiently and consistently, following best practices for architecture and infrastructure.

Overview: The J.A.R.V.I.S. Speech API is a Java-based speech recognition and synthesis API designed to be simple and efficient, utilizing Google's speech engines for functionality. It includes a recognizer, synthesizer, and microphone capture utility, offering a complete, modern, and fully functional speech API in Java. Key features: - Speech recognition and synthesis capabilities - Utilizes Google's speech engines - Complete, modern, and fully functional speech API in Java - Internet connection required for synthesizer and recognizer Tech stack: Java Why it matters: The J.A.R.V.I.S. Speech API provides a straightforward and efficient solution for speech recognition and synthesis, leveraging Google's powerful speech engines. This API is particularly useful for developers looking to integrate speech functionality into their Java projects, offering a comprehensive and modern speech API without the need for external dependencies.

Overview: Twick is an AI-powered video editor SDK built with TypeScript, designed for developers to build custom video apps. It features a timeline editor, drag-and-drop editing, AI captions, and serverless MP4 export. Key features: - AI caption generation - React timeline editor - Canvas-based video editing - GPU-accelerated GL effects - Client-side rendering - Serverless MP4 export Tech stack: TypeScript Why it matters: Twick simplifies video editing for developers, providing a robust, AI-driven solution with real-time caption generation and efficient MP4 export, making it ideal for building custom video applications.

Overview: TernausNetV2 is a fully convolutional network designed for instance segmentation in building detection tasks, achieving second place in the CVPR 2018 DeepGlobe Building Extraction Challenge. Key features: - Fully convolutional network architecture - Uses WideResnet 38 with In-Place Activated BatchNorm - Extends input channels to 11 for better performance - Predicts both binary masks and touching area masks for instance segmentation - Achieved high accuracy on public and private leaderboard Tech stack: Jupyter Notebook Why it matters: TernausNetV2 provides a robust solution for automatic building detection in urban areas, offering significant improvements over previous methods in the CVPR 2018 DeepGlobe Building Extraction Challenge. Repository: ternaus/TernausNetV2 GitHub stars: 544 GitHub description: TernausNetV2: Fully Convolutional Network for Instance Segmentation

### Overview: `wdoc` is a powerful RAG (Retrieval-Augmented Generation) system designed to summarize, search, and query documents across various file types. It's particularly useful for researchers, students, and professionals dealing with extensive information sources. ### Key Features: - **Docker Web UI**: Easy deployment with a Gradio-based web interface for simplified document processing without CLI interaction. - **High Recall and Specificity**: Finds a lot of documents using carefully designed embedding search and aggregates answers using semantic batch to produce a single answer that mentions the source pointing to the exact portion of the source document. - **Supports Virtually Any LLM Provider**: Supports local LLMs and even extra layers of security for super secret stuff. - **Supports Any Filetype**: Queries from all file types (15+ are already implemented). - **Actually Useful AI Powered Summary**: Provides the thought process of the author instead of nebulous takeaways. - **Actually Useful AI Powered Queries**: Provides the sourced indented markdown answer to your questions instead of hallucinated nonsense. - **Extensible**: A tool and a library, turned into an Open-WebUI Tool and available as a Docker web UI for easy deployment. - **Web Search**: Preliminary web search support. ### Tech Stack: Python ### Why It Matters: `wdoc` is designed to solve RAG problems efficiently, making it useful for researchers, students, and professionals dealing with extensive information sources. It supports a wide range of file types and can find a lot of documents using carefully designed embedding search and aggregate

Overview: This project utilizes transformers (based on Hugging Face Transformers) for multi-modal multi-task learning, offering a simplified approach to implementing complex MTL algorithms and extending them to multi-modal scenarios. Key features: - Problem Sampling Strategy - Loss Combination Strategy - Gradient Surgery - Model After Base Model (transformers) Tech stack: Jupyter Notebook Why it matters: This project simplifies the implementation of multi-task learning models, making it straightforward and accessible for developers, especially those working in fields like NLP, CV, and recommendation systems. It also extends the capabilities of MTL to multi-modal scenarios, providing a comprehensive solution for complex data preprocessing, task management, and interaction.

Overview: pygmo is a Python library designed for performing parallel computations of optimization tasks, both global and local, using the asynchronous generalized island model. Key features: - Supports multiple optimization algorithms - Handles various types of optimization problems - Enables easy deployment in massively parallel environments - Provides a unified interface for optimization tasks and problems Tech stack: Python Why it matters: pygmo simplifies the process of implementing and deploying parallel optimization algorithms, making it an essential tool for researchers, educators, and practitioners in fields such as optimization, machine learning, and computational science.

### Overview: StyleSwin is an innovative transformer-based Generative Adversarial Network (GAN) designed for high-resolution image generation. This project addresses the limitations of traditional Convolutional Neural Networks (CNNs) in high-resolution image synthesis by leveraging the strengths of transformers. ### Key Features: - **Transformer-based Architecture:** StyleSwin employs Swin transformers for efficient and expressive image generation. - **Double Attention Mechanism:** This mechanism enhances the model's ability to capture context from both local and shifted windows, leading to improved generation quality. - **Wavelet Discriminator:** To mitigate blocking artifacts during high-resolution synthesis, StyleSwin incorporates a wavelet discriminator to examine spectral discrepancies effectively. ### Tech Stack: - **Language:** Python - **GitHub Stars:** 543 ### Why It Matters: Despite the success of transformers in various vision tasks, this project demonstrates that transformers can achieve par performance in high-resolution image generation, particularly on datasets like CelebA-HQ and FFHQ. The use of Swin transformers and double attention mechanisms significantly improves the model's ability to generate high-resolution images efficiently and effectively.