Projects

### Overview: Open WebUI Tools is a comprehensive toolkit designed to enhance your Open WebUI experience, offering specialized tools, function pipelines, and filters for academic research, image generation, music creation, and more. ### Key Features: - **Tools**: Over 20 specialized tools including arXiv Search, Perplexica Search, Pexels Media Search, YouTube Search & Embed, Xquik X Data Tool, and more. - **Function Pipes**: Advanced tools like Planner Agent v3, arXiv Research MCTS, Multi Model Conversations v2, Resume Analyzer, and more. - **Filters**: Enhance your conversations with Doodle Paint, Prompt Enhancer, Semantic Router, Full Document, and Clean Thinking Tags. ### Tech Stack: Python ### Why It Matters: This toolkit transforms your Open WebUI instance into a powerful AI workstation, supporting academic research, image generation, music creation, and more. Repository: Haervwe/open-webui-tools GitHub Stars: 824

Overview: BambooAI is an AI-driven data analysis environment that leverages Python to provide a transparent and reproducible workflow. It allows analysts to iteratively run Python cells, inspect previous outputs, and verify the evidence behind their reports. Key features: - Iterative analysis - Persistent Python workspace - Executable computational record - Supports local open-weight models Tech stack: Python Why it matters: BambooAI ensures that the analytical process is visible, allowing for easy tracking of numerical results and computational steps. This makes the analysis transparent and interpretable, enhancing trust and reproducibility in the results. Repository: pgalko/BambooAI GitHub stars: 822 GitHub description: BambooAI is an LLM-driven data analysis environment built around a persistent Python kernel, an explicit analysis protocol, and a fully inspectable computational record.

Overview: Adan is an adaptive Nesterov momentum algorithm designed for faster optimization of deep models, as detailed in the paper "Adan: Adaptive Nesterov Momentum Algorithm for Faster Optimizing Deep Models" by Xie et al. This implementation is available on GitHub, where it has received 824 GitHub stars, making it a valuable tool for accelerating the training of deep learning models. Key features: - Adan is an adaptive Nesterov momentum algorithm specifically designed for deep model optimization. - It includes a Riemannian version of Adan (Radan) that is integrated into the Manify library. - Adan is supported in the NeMo framework from NVIDIA. - It is used as the default optimizer in several projects, including the High-Fidelity Text-to-3D Generation Project, Masked Diffusion Transformer V2, D-Adaptation, and more. - Adan is supported in Timm, Paddle, and OpenMMLab projects. - It is also used in the DreamFusion Project and MMClassification. Tech stack: Python Why it matters: Adan is a significant improvement over traditional momentum algorithms, offering faster convergence and better performance on deep learning models. Its adaptive nature makes it particularly effective in scenarios where traditional momentum methods may struggle, making it a valuable tool for researchers and practitioners in the field of deep learning optimization. Repository: sail-sg/Adan Primary language: Python GitHub stars: 824 GitHub description: Adan: Adaptive Nesterov Momentum Algorithm for Faster Optimizing Deep Models

Overview: This repository contains example code for the FLAME 3D head model, demonstrating how to sample from the model, fit it to 3D keypoints, and fit it to 3D scans. Key features: - Sampling: Load and evaluate FLAME model for random parameters - Landmark fitting: Fit FLAME to 3D landmarks - Scan fitting: Fit FLAME to a 3D scan Tech stack: Python Why it matters: This code provides a practical example for researchers and developers interested in using the FLAME model for 3D head fitting and expression analysis.

Overview: This repository is a collection of awesome photogrammetry projects, applications, tools, and resources. It aims to provide a comprehensive list of high-quality photogrammetry-related projects, ensuring that users can find the tools they need for various applications in surveying, mapping, and high-precision metrology. Key features: - A wide range of photogrammetry software, including commercial and open-source options. - Comprehensive datasets and benchmarks for testing and validation. - Information on journals, conference proceedings, books, and introductory papers. - Courses and tutorials for learning photogrammetry. - Scientific conferences and related awesome lists for staying updated in the field. Tech stack: The repository uses Markdown and GitHub to organize and share information. Why it matters: This repository serves as a valuable resource for anyone interested in photogrammetry, providing a platform to discover and explore the latest tools, datasets, and educational materials in the field. Repository: awesome-photogrammetry/awesome-photogrammetry GitHub stars: 822

### Overview: This repository is a curated list of awesome scene understanding papers, inspired by the popular [awesome-computer-vision](https://github.com/jbhuang0604/awesome-computer-vision) repository. It aims to provide a comprehensive overview of the latest research in scene understanding, covering various techniques and datasets used in the field. ### Key Features: - 📷 Multi-view images - 🎲 Point cloud - 🌟 3D models and datasets - 🌟 Survey articles and datasets - 🌟 Workshops and tutorials - 🌟 Realistic dataset repositories ### Tech Stack: The repository is written in Python and uses Markdown for documentation. ### Why It Matters: This list serves as a valuable resource for researchers, developers, and enthusiasts interested in advancing the field of scene understanding. It provides a curated collection of papers, datasets, and resources that can help in understanding and improving the techniques used in scene recognition and understanding. ### GitHub Stars: The repository has received 823 GitHub stars, indicating its popularity and usefulness among the developer community. ### Related Resources: - [3D Machine Learning](https://github.com/timzhang642/3D-Machine-Learning) - [Awesome Holistic 3D](https://github.com/holistic-3d/awesome-holistic-3d) - [Awesome Planar Reconstruction](https://github.com/chenzhaiyu/awesome-planar-reconstruction) - [Wireframe](https://github.com/Delay-Xili

### Overview: This repository is a curated list of 3D Vision papers related to the Robotics domain, focusing on the era of large models, inspired by the awesome-computer-vision project. It includes papers, codes, and related websites. ### Key features: - **Policy Learning**: Papers and implementations related to policy learning in robotics. - **Pretraining**: Papers and implementations related to pretraining models for robotics. - **VLM and LLM**: Papers and implementations related to Vision Language Models and Large Language Models for robotics. - **Representations**: Papers and implementations related to 3D representations and visual foundation models. - **Simulations, Datasets, and Benchmarks**: Papers and implementations related to simulations, datasets, and benchmarks for robotics. - **Citation**: Papers and implementations related to citation and references. ### Tech stack: Python, Jupyter Notebooks, and various machine learning libraries. ### Why it matters: This repository provides a comprehensive overview of recent advancements in 3D vision and robotics, making it a valuable resource for researchers and developers in the field.

Overview: This repository contains CUDA/C++ code for fusing multiple depth maps into a projective truncated signed distance function (TSDF) voxel volume, which can be used to generate high-quality 3D surface meshes and point clouds. Key features: - Fusion of multiple depth maps into a TSDF voxel volume - Generation of 3D surface point clouds and 3D surface meshes - Support for Ubuntu 14.04 and 16.04 - Python version available for both CPU and GPU support Tech stack: CUDA/C++ Why it matters: This code is essential for creating accurate 3D models from multiple depth maps, making it valuable for applications in computer vision, robotics, and 3D modeling.

Overview: Kur is a system designed to simplify the creation and application of state-of-the-art deep learning models, making it accessible to both beginners and advanced users alike. Key features: - Descriptive specification files for easy model design and understanding. - Supports Theano, TensorFlow, and PyTorch. - Multi-GPU support out-of-the-box. - Jinja2 templating engine for model exploration. - Installation via pip or cloning the repository. Tech stack: Python 3.4 or greater. Why it matters: Kur simplifies the process of building and applying deep learning models, making it easier for everyone involved in the machine learning community, from beginners to experts.

Overview: Helix is a private agent fleet that allows developers to run Claude, Codex, Gemini, and open models on their own infrastructure. Each agent has its own GPU-accelerated desktop, and the fleet is managed on a shared Kanban board with a full private AI Stack. Key features: - Full desktop per agent - Fleet visibility - Multiplayer collaboration - High-density isolation Tech stack: Go Why it matters: Helix provides a private, isolated environment for running AI agents, allowing developers to collaborate and review changes in a more efficient and productive manner.

### Overview: Xyne Spaces is a collaborative platform designed for organizations, providing a unified context layer for humans and agents. It integrates with existing tools like Slack, Google Workspace, and Microsoft 365, normalizing and serving context through permission-aware APIs, ensuring that access control respects the permissions of the underlying work. ### Key Features: - **Context Store**: Normalizes and indexes context from various tools, serving it through permission-aware APIs. - **Collaboration**: Real-time work with team chat, calls, and collaboration tools. - **Agentic Workflows**: Agents can search, read, and act on context through the same permission model. - **Apps**: Customer Support Desk, Tickets, and other productivity apps built around the context layer. ### Tech Stack: TypeScript ### Why It Matters: Xyne Spaces provides a single, normalized context layer that respects access control, ensuring that agents and people can only access the context they are authorized to see. This centralizes context, making it easier to search, automate, and collaborate across different tools and platforms. ### Repository: juspay/xyne-spaces ### GitHub Stars: 822 ### GitHub Description: The AI Org-OS — a collaborative platform for humans and agents. Explore: https://spaces.xyne.ai

Overview: Sentinel is a real-time monitoring system designed for local area networks (LAN). It utilizes Android devices as network cameras, enabling distributed image acquisition, real-time video streaming, and AI-driven monitoring and analysis. The system is built on a layered architecture of mobile capture, PC processing, and browser control, supporting features such as live preview, local video recording, and structured event output for data acquisition and analysis. Key features: - Real-time video preview - System parameter customization - AI-triggered monitoring and controllable cognitive output - Mobile camera capture app (CamFlow) Tech stack: Python Why it matters: Sentinel is a lightweight, efficient solution for local monitoring and data acquisition, suitable for both individual use and as an engineering prototype for visual data analysis and intelligent monitoring.

### Overview: AutoMem is a long-term memory system designed for AI assistants, providing access to the decisions, preferences, notes, and context that matter. It helps AI assistants recall the "why" behind their actions, rather than just the words they recognize. ### Key features: - **Graph + Vector Store**: AutoMem uses a graph and vector store to recall decisions, relationships, and context across sessions. - **Private Memory**: AutoMem stores memories locally, ensuring privacy and avoiding dependency on external services. - **Efficient Retrieval**: AutoMem retrieves memories directly through its graph and vector index, keeping normal retrieval fast and avoiding an extra generative-LLM charge. ### Tech stack: Python ### Why it matters: AutoMem provides a more comprehensive and efficient way to remember and recall context for AI assistants, making them more useful and efficient in their interactions.

### Overview: MonoScene: Monocular 3D Semantic Scene Completion ### Key features: - Monocular 3D Semantic Scene Completion - 3D Semantic Occupancy Prediction from a single image - Generalized 3D Occupancy Prediction - Urban 3D Panoptic Scene Completion with Uncertainty Awareness ### Tech stack: Python ### Why it matters: This project introduces a unified approach for 3D occupancy prediction, enabling generalized and unconstrained urban 3D scene completion from a single image. The method is evaluated on SemanticKITTI and KITTI-360 datasets, demonstrating superior performance compared to existing methods. The project also includes a tutorial on defining viewpoints in mayavi and a guide on installing mayavi with Python 3 on Ubuntu 20.04 using pip or conda.

Overview: `bigarrow` is a macOS command-line tool that allows AI agents to draw arrows, boxes, and text on any window, making it easy to interact with applications without needing to click on buttons. It works seamlessly with Claude Code and Codex, and is MIT licensed. Key features: - Draws arrows, boxes, and text on any window - Clicks through without changing focus - Supports macOS and iOS - MIT licensed Tech stack: Swift binary Why it matters: Simplifies interaction with applications without requiring user intervention, making it useful for developers and users alike. Repository: franzenzenhofer/big-arrow-on-the-screen GitHub stars: 814 GitHub description: A macOS command-line tool for drawing arrows, boxes, and text on any window, making it easy to interact with applications without needing to click on buttons. Supports Claude Code and Codex, and is MIT licensed.

Overview: Caer is a lightweight, high-performance Vision library designed for high-performance AI research. It simplifies Computer Vision by abstracting away unnecessary boilerplate code, making it ideal for students, researchers, hobbyists, and experts in Deep Learning and Computer Vision. Key features: - GPU-accelerated Computer Vision operations - Type-checked API for ease of use - Standard test images and example data - Flexible and well-integrated with other libraries Tech stack: Python Why it matters: Caer provides a modern, type-checked approach to Computer Vision, making it accessible and efficient for various research and development needs. Repository: jasmcaus/caer Primary language: Python GitHub stars: 822 GitHub description: High-performance Vision library in Python. Scale your research, not boilerplate. README excerpt: <!--  --> Key features: - GPU-accelerated Computer Vision operations - Type-checked API for ease of use - Standard test images and example data - Flexible and well-integrated with other libraries Tech stack: Python Why it matters: Caer provides a modern, type-checked approach to Computer Vision, making it accessible and efficient for various research and development needs.

Overview: WorldSeed is a multi-agent world engine that simulates emergent outcomes through defined rules, agents, and consequences. It allows for the creation of various worlds, from research communities to AI tool pilots to AI layoffs scenarios. Key features: - Scene-agnostic engine that can run any world definition. - Rules, agents, and consequences define the world. - Agents interact until useful artifacts emerge. - Supports production rooms, simulations, games, and fictional worlds. - Demo showcases the engine's capabilities, including an example of AI tool pilot lab and an AI layoffs scenario. Tech stack: Python. Why it matters: WorldSeed provides a framework for exploring and simulating different scenarios through agent-based modeling, which can be applied in various fields such as research, AI development, and organizational behavior studies. Repository: AIScientists-Dev/WorldSeed Primary language: Python GitHub stars: 821 GitHub description: More is Different. A multi-agent world engine where AI agents live, talk, compete, ally.

### Overview: Artemis is an open-source platform designed to provide interactive learning experiences with individual feedback, suitable for universities to deploy on their own infrastructure. It offers scalable, interactive learning with features like automated feedback, adaptive learning, and responsible AI support. ### Key features: - **Interactive Learning with Individual Feedback**: Artemis enables universities to offer interactive learning experiences with individual feedback. - **Programming, Modeling, Text, Quiz, and File-Upload Exercises**: Artemis supports a variety of exercise types, including programming, modeling, text, quiz, and file-upload exercises. - **Authentic Assessment and Online Exams**: Artemis uses the same exercise types for coursework and examinations, with exercise variants per student and plagiarism checks. - **Adaptive Learning**: Artemis adapts learning paths and competencies to individual progress. - **Responsible AI Support**: Artemis includes optional integrations like Iris, Athena, and Hyperion, which are configuration-gated and leave instructors and tutors responsible for assessment decisions. - **Open Digital Infrastructure**: Artemis is MIT-licensed, developed in public, self-hosted, and designed to scale horizontally for large courses. ### Tech stack: Java ### Why it matters: Artemis is essential for universities looking to enhance their learning experiences with interactive, individual feedback, while maintaining a responsible and transparent approach to AI integration.

Overview: This project implements a sparse mixture of experts language model inspired by Andrej Karpathy's "makemore" project, focusing on a scratch implementation using PyTorch. Key features: - Sparse mixture of experts architecture - Top-k and noisy top-k gating mechanisms - Kaiming He initialization - Expert capacity update Tech stack: PyTorch Why it matters: This implementation provides a detailed guide on the sparse mixture of experts architecture, focusing on readability and hackability, making it useful for developers looking to understand and improve the model.

### Overview: This repository provides a simple tutorial for AI researchers to use Low-Rank Adaptation (LoRA) within the Diffusers framework. It offers a straightforward way to play with various Lora models from active communities like Hugging Face and Civitai, including models from the Stable Diffusion webUI, and supports the ControlNet and T2I-Adapter for Diffusers. ### Key Features: - Load or insert pre-trained Lora weights into Diffusers framework. - Support models in the safetensors format, which is not well compatible with Diffusers. - Includes examples for both full models and Lora-only models. - Provides a script to convert checkpoint files to Diffusers format. ### Tech Stack: Python ### Why it Matters: This tutorial simplifies the process of integrating Lora into Diffusers, making it accessible for AI researchers without requiring extensive knowledge of the underlying frameworks. It allows for efficient fine-tuning of models on large base models, significantly reducing computational resources.