Projects

Overview: SplaTAM is a state-of-the-art dense RGB-D SLAM algorithm that efficiently tracks and maps 3D Gaussians for real-time applications. It has been presented at CVPR 2024 and is designed for real-time dense SLAM tasks. Key features: - Real-time dense SLAM with efficient 3D Gaussian tracking - GPU-accelerated for fast performance - Supports multiple LiDAR sensors - Integrates with popular RGB-D cameras Tech stack: Python, PyTorch, CUDA Why it matters: SplaTAM offers a significant improvement in real-time dense SLAM by efficiently tracking and mapping 3D Gaussians, making it ideal for applications requiring real-time performance and accuracy. Repository: spla-tam/SplaTAM GitHub stars: 2214 GitHub description: SplaTAM: Splat, Track & Map 3D Gaussians for Dense RGB-D SLAM (CVPR 2024) README excerpt: <!-- PROJECT LOGO --> <p align="center"> <h1 align="center">SplaTAM: Splat, Track & Map 3D Gaussians for Dense RGB-D SLAM</h1> <h3 align="center">CVPR 2024</h3> <p align="center"> <a href="https://nik-v9.github.io/"><strong>Nikhil Keetha</strong></a> · <a href="https://jaykarhade.github

### Overview: Rubix ML is a high-level machine learning and deep learning library for PHP, providing a developer-friendly API with over 40 supervised and unsupervised learning algorithms, support for ETL, preprocessing, and cross-validation, and open-source availability. ### Key Features: - **Developer-Friendly API**: Easy to use and integrate with existing PHP projects. - **40+ Algorithms**: A wide range of supervised and unsupervised learning methods. - **ETL, Preprocessing, and Cross-Validation Support**: Tools for data preparation and model evaluation. - **Open Source and Commercially Free**: Free to use, with optional extensions for better performance. ### Tech Stack: PHP ### Why It Matters: Rubix ML simplifies the process of building machine learning models using PHP, making it accessible to developers without extensive knowledge in machine learning. The library's comprehensive set of algorithms and features make it suitable for various applications, from image recognition to credit risk prediction, making it a valuable tool for developers looking to leverage machine learning in their projects.

Overview: PicoLM is a pure C implementation of a 1-billion parameter language model, designed for running on a $10 board with 256MB RAM. It eliminates the need for the internet, cloud, or Python dependencies, making it a perfect local AI assistant. Key features: - Pure C implementation - No dependencies - Single binary file - 45MB RAM usage - Runs on a $10 board Tech stack: C Why it matters: PicoLM provides a fully offline AI assistant that can be used for various applications, including local chatbots and AI-driven tools, without the need for internet connectivity or cloud services.

Overview: The Operator Splitting QP Solver is a numerical optimization package designed to solve quadratic programming problems with linear constraints. Key features: - Solves quadratic programming problems with linear constraints. - Supports both continuous and bounded variables. - Utilizes the operator splitting method for efficient computation. Tech stack: C Why it matters: The OSQP solver is highly efficient and effective for solving optimization problems in various applications, including control theory, machine learning, and signal processing. Repository: osqp/osqp GitHub stars: 2212 GitHub description: The Operator Splitting QP Solver

### Overview: 本项目是一个面向 OpenClaw 的完整学习教程，帮助你从零开始掌握这个强大的命令行 AI 助理系统。无论你是想快速上手使用 OpenClaw 提升效率，还是想深入理解其原理并构建自己的版本，本教程都能为你提供清晰的学习路径。 ### Key features: - 领养龙虾（使用篇）：11 章 + 7 个附录，安装（Ch1-3）+ 核心配置（Ch4-6）+ 扩展运维（Ch7-9）+ 安全与客户端（Ch10-11） - 龙虾大学（场景实战篇）：围绕 Skills 选型与典型工作流，给出可直接复用的实战案例 - 构建龙虾（开发篇）：11 章，先拆解 OpenClaw 源码与替代方案，再进入 Skills、渠道和完整定制 ### Tech stack: Python ### Why it matters: 本项目提供了一个全面的学习路径，帮助用户从零开始掌握 OpenClaw 的技能系统和自动化能力，适用于零基础用户、效率达人、技术爱好者和开发者。 ### Repository: datawhalechina/hello-claw ### GitHub stars: 2213 ### GitHub description: 哈喽！龙虾 🙋‍♀️ Adopt from scratch and build your first claw 🦞 来领养你的第一只龙虾！

### Overview: HRM-Text is a 1B text generation model based on the HRM architecture, designed to be highly efficient and accessible. It leverages task completion and latent space reasoning to enhance pretraining, reducing compute and data requirements significantly. ### Key Features: - **1B Parameter Model**: HRM-Text is a 1B parameter model, offering a significant reduction in computational and data requirements compared to traditional pretraining methods. - **Hierarchical Recurrent Architecture**: Built on a hierarchical recurrent architecture, PrefixLM sequence packing, FlashAttention 3 kernels, PyTorch FSDP2 training, and evaluation tools. - **Task Completion and Latent Space Reasoning**: Strengthened by task completion and latent space reasoning, HRM-Text aims to improve the quality and relevance of generated text. ### Tech Stack: - Python ### Why it Matters: - **Efficiency**: HRM-Text requires only 130-600x less compute and 150-900x less data compared to traditional pretraining methods. - **Reduced Costs**: The model is designed to be pre-trained for approximately $1000, making it highly accessible and cost-effective. - **Improved Quality**: By leveraging task completion and latent space reasoning, HRM-Text aims to generate more contextually relevant and task-specific text. ### GitHub Stars: 2203 ### Primary Language: Python ### GitHub Description: HRM-Text is a 1B text generation model based on the HRM architecture, strengthened by task

Overview: xLSTM is an advanced Recurrent Neural Network architecture designed to enhance language modeling capabilities, surpassing traditional LSTM and Transformer models. It introduces exponential gating, normalization, and matrix memory to improve performance and efficiency. Key features: - xLSTM Large: A 7B parameter language model optimized for large-scale inference. - xLSTM 7B: A 7B parameter language model specifically optimized for fast and efficient inference. - xLSTM 7B poster: A high-quality visual representation of the xLSTM architecture. - xLSTM Large 7B implementation: A standalone single file implementation of the xLSTM Large architecture. Tech stack: Python, PyTorch, Hugging Face. Why it matters: xLSTM offers significant improvements in language modeling performance, particularly in terms of training throughput and stability, making it a valuable tool for researchers and developers in the field of natural language processing. Repository: NX-AI/xlstm GitHub stars: 2211 Primary language: Python GitHub description: Official repository of the xLSTM.

### Overview: llmwiki is a knowledge compiler that transforms raw sources into an interlinked, citation-traceable markdown wiki. Inspired by Karpathy's LLM Wiki pattern, it simplifies the process of compiling and organizing knowledge. ### Key Features: - **Wiki Behavior Separation:** Compilation, retrieval, and review now describe what they need from storage instead of touching files directly. - **Converse Gateways:** Use a bearer-authenticated Converse endpoint as your provider. - **Pages That Fit the Source:** `llmwiki compile --no-pages-limit` drops the fixed concept count. - **A Viewer That Keeps Up:** `llmwiki view` refreshes after a compile without a restart and reads its port from `LLMWIKI_VIEW_PORT`. ### Tech Stack: TypeScript ### Why It Matters: llmwiki simplifies the process of compiling and organizing knowledge, making it easier for both agents and humans to browse, query, lint, export, and reuse knowledge. The default profile preserves the classic concepts-and-queries layout, while optional profiles add domain-specific types and workflows without adding domain branches to the compiler.

### Overview: IMS Toucan is a toolkit for training, using, and teaching state-of-the-art Text-to-Speech Synthesis, developed at the Institute for Natural Language Processing (IMS), University of Stuttgart, Germany. ### Key Features: - **Controllable and Fast**: IMS Toucan provides a controllable and fast Text-to-Speech synthesis system. - **Supports 7,000+ Languages**: It supports over 7,000 languages, making it versatile for various applications. - **Training and Inference**: The toolkit supports both training and inference, with a focus on speed and efficiency. ### Tech Stack: Python ### Why it Matters: - **Efficiency**: The system is designed to be efficient, making it suitable for real-time applications. - **Versatility**: Its support for over 7,000 languages makes it highly adaptable to various linguistic needs. - **Speed**: The toolkit is designed to be fast, reducing the time required for text-to-speech conversion. ### Repository: DigitalPhonetics/IMS-Toucan - **Stars**: 2210 - **Description**: Controllable and fast Text-to-Speech for over 7,000 languages! - **GitHub Release**: 2.0.0 - **GitHub Downloads**: 10,000 - **GitHub Stars**: 2210 - **GitHub Description**: Controllable and fast Text-to-Speech for over 7,000 languages! ### Installation: - **Basic Requirements**:

### Overview: The A2A Python SDK is a comprehensive library designed for building agentic applications that adhere to the Agent2Agent (A2A) Protocol. It offers a modern, asynchronous architecture and supports various communication protocols and database backends, ensuring high performance and flexibility. ### Key Features: - **A2A Protocol Compliant:** Ensures that applications are built according to the A2A Protocol. - **Extensible:** Easily integrates with different communication protocols and database backends. - **Asynchronous:** Built on modern async Python for efficient performance. - **Optional Integrations:** Provides support for HTTP servers (FastAPI, Starlette), gRPC, OpenTelemetry for tracing, and SQL databases (PostgreSQL, MySQL, SQLite). ### Tech Stack: - Python 3.10+ - `uv` (recommended) or `pip` ### Why It Matters: This SDK is essential for developers looking to build agentic applications that follow the Agent2Agent (A2A) Protocol. It ensures compliance with the protocol, supports various integrations, and leverages modern Python features for enhanced performance and flexibility. ### Repository: The repository is hosted on GitHub at [a2aproject/a2a-python](https://github.com/a2aproject/a2a-python). ### GitHub Stars: The repository has been starred by 2,211 developers, indicating its popularity and utility in the community. ### Compatibility: The SDK supports the following compatibility modes: - **`1.0`** (latest version): JSON-RPC, HTTP+JSON/REST

### Overview: This repository is a collection of advice for prospective and current PhD students in computer science and related fields. It aims to provide valuable guidance on various aspects of the PhD application and research process. ### Key Features: - General advice on applying to PhD programs. - Tips for writing a compelling statement of purpose. - Advice on visiting schools and choosing advisors. - Survival guides for PhD students. - Research advice, including reading and writing. - Writing advice, including blog posts and reviewing. - Presenting advice, including advising students. - Teaching advice, including networking. - Fellowship application advice. - Networking advice. - Organizing workshops and tutorials. - Attending academic conferences. - Job search advice. - Memoirs and reflections. ### Tech Stack: None specified. ### Why it Matters: This collection of advice is invaluable for both prospective and current PhD students, providing comprehensive guidance on navigating the PhD application and research process. It serves as a valuable resource for anyone seeking to succeed in the highly competitive field of computer science and related disciplines.

Overview: MOSS-Transcribe-Diarize is a 0.9B model designed for long-form transcription in 50+ languages, providing speaker diarization, timestamps, and acoustic event awareness. It is part of the MOSS-Transcribe-Diarize family and is available on Hugging Face. Key features: - Speaker diarization - Timestamps - Acoustic event awareness Tech stack: Python Why it matters: MOSS-Transcribe-Diarize is a state-of-the-art model for accurately transcribing long-form audio with speaker labels and timestamps, making it essential for applications requiring precise transcription and speaker identification in diverse languages. Repository: OpenMOSS/MOSS-Transcribe-Diarize Primary language: Python GitHub stars: 2180 GitHub description: A 0.9B model for long-form transcription in 50+ languages with speaker diarization, timestamps, and acoustic event awareness.

Overview: Unsupervised Data Augmentation (UDA) is a semi-supervised learning method that significantly reduces the need for labeled examples, achieving state-of-the-art results across various tasks such as text and image classification. Key features: - Improved performance on tasks with limited labeled data - Reduced error rate compared to state-of-the-art methods - Efficient use of unlabeled data through back translation augmentation Tech stack: Python, Tensorflow, Google Cloud TPU Why it matters: UDA enables more effective use of limited labeled data, leading to significant improvements in accuracy and efficiency in semi-supervised learning tasks.

### Overview: Lion is a new optimizer discovered by Google Brain using genetic algorithms that is purportedly better than Adam(w) in Pytorch. It is designed to be accessible and used as soon as possible by everyone to train great models. ### Key Features: - **Learning Rate and Weight Decay**: The authors suggest a learning rate that is 3-10 times smaller than Adam's, with a decoupled weight decay that is 3-10 times larger. The initial, peak, and end values of the learning rate schedule should be changed simultaneously with the same ratio as AdamW. - **Learning Rate Schedule**: Uses the same learning rate schedule as AdamW but observes better gains with a cosine decay schedule for ViT training. - **β1 and β2**: The default values for β1 and β2 in AdamW are 0.9 and 0.999, respectively, while in Lion, they are discovered through program search and set as 0.9 and 0.99, respectively. Using `β1=0.95` and `β2=0.98` can help mitigate instability during training. - **Updates**: - Seems to work for local English Wik8 language modeling. - Experiment results show Lion with a 3x smaller learning rate beats Adam. - Using a 10x smaller learning rate rule of thumb from the paper led to the worst run. - Positive results for language modeling and significant text-to-image training, with negative results for problems and architectures outside the

Overview: Beelzebub is an open-source deception framework that leverages AI for system virtualization, enabling security teams to study and analyze activity directed at decoy services. Key features: - Define services, routes, and response rules in YAML. - Use static handlers for predictable behavior or LLM-powered responses for adaptive interactions. - Compile trusted Go plugins to extend services and responses. - Capture evidence from decoy interactions, including commands, HTTP request details, and session context. - Run the framework independently or connect it to Beelzebub Platform for further investigation. Tech stack: Go Why it matters: Beelzebub provides a configurable way to study activity directed at decoy services, allowing security teams to validate and analyze interactions, and to support evidence-based investigation and workflow integration. Repository: beelzebub-labs/beelzebub GitHub stars: 2203 GitHub description: A secure low code deception runtime framework, leveraging AI for System Virtualization.

Overview: Skales is a personal AI agent designed for macOS, Windows, Linux, Android, and iOS. It automates coding, desktop, and browser automation, and schedules tasks. Skales integrates with 26 AI providers, supports BYOK, and runs without Docker or a terminal. Key features: - Skales Code for coding tasks - Desktop and browser automation - Scheduled tasks - Teams collaboration - Agent2Agent, MCP, and computer use - Voice integration - Local LLMs (Ollama) - Agent Skills (SKILL.md) - 15+ AI providers - Bring your own key from 15+ providers or run fully offline Tech stack: Skales is built on a combination of AI models and a lightweight, desktop-based architecture. Why it matters: Skales offers a powerful, low-maintenance solution for automating tasks and managing workflows on your own computer, without the need for cloud-based services. Its integration with multiple AI providers makes it versatile and customizable, suitable for both developers and non-developers.

Overview: This is an open-source Chinese LLaMA2 7B model and instruction dataset that can be downloaded and run. Key features: - Downloadable and executable Chinese LLaMA2 7B model - Commercially supported - Compatible with all LLaMA2 Chat model optimizations - Includes Chinese-English SFT dataset with 1 million instructions Tech stack: Python Why it matters: This model and dataset provide a practical and accessible way to work with Chinese LLaMA2, making it easier for developers to integrate and utilize the model in their projects.

Overview: The AI Compendium, formerly known as the ML Compendium, is a comprehensive resource for machine learning and deep learning, covering a wide array of topics from modern algorithms to specialized areas like NLP and computer vision. It is designed to help learners of all levels connect with influential authors and support the community. Key features: - A wide range of topics, including modern machine learning algorithms and deep learning techniques. - Specialized areas such as NLP, audio processing, computer vision (classic and deep), time-series analysis, anomaly detection, and graphs. - Strategic themes like data science management, team building, and practical essentials like product management, design, and technology stacks from a data science perspective. - A streamlined path to valuable resources and authors. Tech stack: The book is maintained on GitHub. Why it matters: The AI Compendium serves as a go-to resource for learners of all levels, providing a streamlined path to invaluable authors and resources. It is designed to save learners time by filtering through content and connecting them with the community. Repository: orico/www.mlcompendium.com GitHub stars: 2200 GitHub description: The AI Compendium, formerly the Machine & Deep Learning Compendium, curates resources and explanations of machine learning (ML) and deep learning, from a business problem through data, AI, engineering, and product.

Overview: This repository contains the PyTorch implementation of two state-of-the-art vision transformer-based models for human pose estimation: ViTPose and ViTPose++. These models were presented at NeurIPS 2022 and TPAMI 2023, respectively, and achieved impressive results on the MS COCO test-dev set. Key features: - ViTPose: Simple Vision Transformer Baselines for Human Pose Estimation - ViTPose++: Vision Transformer for Generic Body Pose Estimation Tech stack: Python Why it matters: These models offer a significant improvement in accuracy and efficiency compared to traditional methods, making them highly valuable tools for researchers and practitioners in the field of computer vision and human pose estimation. Repository: ViTAE-Transformer/ViTPose Primary language: Python GitHub stars: 2200 GitHub description: The official repo for NeurIPS'22 "ViTPose: Simple Vision Transformer Baselines for Human Pose Estimation" and TPAMI'23 "ViTPose++: Vision Transformer for Generic Body Pose Estimation".

Overview: This repository serves as a developer asset hub for NVIDIA's Nemotron, providing training recipes, usage cookbooks, datasets, and end-to-end reference examples to help build with Nemotron models. Key features: - Training recipes - Usage cookbooks - Datasets - End-to-end reference examples Tech stack: Jupyter Notebook Why it matters: This repository offers a comprehensive resource for developers working with NVIDIA's Nemotron models, facilitating community innovation, compute efficiency, high accuracy, and flexible deployment options.