Projects

### Overview: **manif** is a small C++11 header-only library designed specifically for robotics applications, providing Lie theory functionalities such as Euclidean space, rotations, rigid motions, and more. It aims to simplify the development of state estimation algorithms by offering a comprehensive yet accessible introduction to Lie theory tailored for robotics. ### Key features: - **Euclidean Space**: ℝ(n) - **Rotations in the Plane**: SO(2) - **Rigid Motion in the Plane**: SE(2) - **Rotations in 3D Space**: SO(3) - **Rigid Motion in 3D Space**: SE(3) - **Extended Pose in 3D Space**: SE_2(3) - **Special Galilean Group**: SGal(3) - **Composite Manifold**: Bundle<> ### Tech stack: C++ ### Why it matters: **manif** simplifies the development of state estimation algorithms in robotics by providing a header-only library with Python 3 wrappers. The comprehensive introduction to Lie theory in the paper ensures that even those without advanced mathematical backgrounds can grasp the foundational concepts effectively. ### Repository: artivis/manif ### GitHub stars: 1826

### Overview: RustQuant is a Rust library designed for quantitative finance. It provides essential tools and functionalities for developing financial applications and models. ### Key Features: - **Algorithmic Adjoint Differentiation (AAD)**: For efficiently computing gradients of scalar output functions. - **Financial Instruments**: Implementations for bonds, options, and money, including their pricing. - **Data Types and Sources**: Data types for pricing and similar tasks, including CSV, JSON, and Parquet file support. - **Error Handling**: A RustQuant error handling module. - **ISO Codes**: Implementations for ISO currency codes, country codes, and market identifier codes. - **Statistical Functions**: Statistical distributions, Fast Fourier Transform, numerical integration, optimization, and risk-reward metrics. - **Financial Term Structures**: Term-structures and surfaces for financial modeling. ### Tech Stack: Rust ### Why It Matters: RustQuant offers a comprehensive set of tools for quantitative finance, making it an essential library for developers working in this field. Its robust features and efficient performance make it suitable for a wide range of applications in quantitative finance. --- Repository: avhz/RustQuant GitHub Stars: 1826

### Overview: ClaudePrism is an offline-first scientific writing workspace powered by Claude, allowing users to compile and edit LaTeX, Python, and 100+ scientific skills locally on their desktop. ### Key Features: - **Local Compilation**: Files are stored locally and compiled offline. - **AI Features**: AI models are only used for inference, not for compiling or editing documents. - **Scientific Skills**: 100+ domain-specific skills, including bioinformatics, cheminformatics, and machine learning. - **Project Setup**: Easy installation and project setup with a template gallery and project wizard. ### Tech Stack: Tauri 2 + Rust ### Why It Matters: ClaudePrism offers a local-first experience for scientific writing, ensuring privacy and control over data. It integrates seamlessly with Python environments and includes a variety of scientific skills, making it suitable for researchers and scientists.

Overview: This repository serves as a practical course for developers, engineers, and researchers, focusing on Large Language Models (LLMs) and their applications. The course covers techniques, libraries, projects, and enterprise solutions, providing a hands-on experience using models from OpenAI and the Hugging Face library. Key features: - Techniques and Libraries - Projects - Enterprise Solutions Tech stack: Jupyter Notebook Why it matters: The course offers a comprehensive guide to building and applying Large Language Models, covering essential techniques, libraries, and real-world projects, making it valuable for developers looking to enhance their skills in this rapidly evolving field.

Overview: NVIDIA NeMo Curator is a scalable data pre-processing and curation toolkit designed for large-scale text, image, video, and audio datasets used in AI training. It helps ML engineers and data teams build repeatable, GPU-accelerated pipelines that can be run on a laptop or across a multi-node Ray cluster. Key features: - Curator on Slurm: Run multi-node Ray pipelines on HPC clusters. - Audio Curation: Build ALM and speech datasets with composite quality filtering, audio tagging, and speaker diarization. - Inference Server: Spin up an OpenAI-compatible LLM endpoint inside your pipeline for SDG, classification, and synthetic data workflows. Tech stack: Python. Why it matters: NeMo Curator simplifies the process of creating and managing pipelines for large datasets, making it easier to build and maintain repeatable workflows that can be scaled up or down as needed. Repository: NVIDIA-NeMo/Curator Primary language: Python GitHub stars: 1814 GitHub description: Scalable data pre processing and curation toolkit for LLMs README excerpt: <div align="center"> <a href="https://github.com/NVIDIA-NeMo/Curator/blob/main/LICENSE"></a> <a href="https://codecov.io/github/NVIDIA-NeMo/Curator">![codecov](https://codecov.io/github/NVIDIA-NeMo/Curator

### Overview: Dive is an open-source MCP Host Desktop Application that seamlessly integrates with any LLMs supporting function calling capabilities. ✨ ### Key features: - 🌐 Universal LLM Support: Compatible with ChatGPT, Anthropic, Ollama, and OpenAI-compatible models. - 💻 Cross-Platform: Available for Windows, MacOS, and Linux. - 🔄 Model Context Protocol: Enables seamless MCP AI agent integration on both stdio and SSE mode. - ☁️ OAP Cloud Integration: One-click access to managed MCP servers via [OAPHub.ai]. - 🏗️ Dual Architecture: Modern Tauri version alongside traditional Electron version for optimal performance. - 🌍 Multi-Language Support: Supports 24+ languages including English, Traditional Chinese, Simplified Chinese, Spanish, Japanese, Korean, German, French, Italian, Portuguese, Russian, Thai, Vietnamese, Filipino, Indonesian, Polish, Turkish, Ukrainian, Swedish, Norwegian, Finnish, and Lao. - ⚙️ Advanced API Management: Multiple API keys and model switching support with `model_settings.json`. - 🛠️ Granular Tool Control: Enable/disable individual MCP tools for precise customization. - 💡 Custom Instructions: Personalized system prompts for tailored AI behavior. - ⌨️ Keyboard Shortcuts: Comprehensive hotkey support for efficient navigation and operations (rename, settings, reload, new chat, etc.). - 📝 Chat Draft Saving: Automatically saves chat input drafts to prevent data loss. - 🔄 Auto-Update Mechanism: Automatically checks

### Overview: PPL Quantization Tool (PPQ) is a powerful offline neural network quantization tool designed for industrial applications, aiming to accelerate neural network inference on low-power and resource-constrained devices. ### Key Features: - Supports multiple quantization formats, including FP8, FP16, and INT8. - Provides a flexible and extensible quantization pipeline that allows users to customize the quantization process. - Integrates with popular inference frameworks like ONNX, TensorRT, and ONNXruntime, enabling seamless integration. - Offers advanced features such as quantization-aware training (QAT) and quantization-aware optimization (QAO). ### Tech Stack: Python ### Why it Matters: By enabling efficient quantization of neural networks, PPQ helps reduce computational complexity and power consumption, making it ideal for edge devices and IoT applications where resource constraints are significant.

Overview: DeTi*k*Zify is a project that synthesizes graphics programs for scientific figures and sketches using TikZ. It aims to streamline the process of creating high-quality, reproducible graphics. Key features: - Automatic TikZ code generation - Integration with LaTeX - Support for various scientific figures and sketches - Easy customization and modification Tech stack: Python Why it matters: DeTi*k*Zify simplifies the creation of scientific graphics, making it easier for researchers and scientists to produce high-quality, reproducible figures without the need for extensive LaTeX knowledge. This tool enhances collaboration and efficiency in scientific research and education.

### Overview: This repository is a curated list of deep learning resources, designed to be useful for both learning and reference. The resources cover a wide range of topics, from online classes and books to practical implementations and other foundational theories. ### Key Features: - **Online Classes:** Includes courses on deep learning and recurrent neural networks, such as the DL&RNN Course by Neuraxio. - **Books:** Lists recommended books for deep learning and machine learning. - **Librairies and Implementations:** Provides resources for popular deep learning libraries and implementations. - **Other Math Theory:** Covers advanced topics like gradient descent algorithms and complex numbers in digital signal processing. - **Papers:** Includes papers on recurrent neural networks, convolutional neural networks, attention mechanisms, and other deep learning topics. - **YouTube and Videos:** Offers tutorials and explanations on various deep learning concepts. - **Miscellaneous:** Includes links to other hubs and resources. ### Tech Stack: Python and Jupyter Notebooks. ### Why it Matters: This list serves as a comprehensive guide for anyone looking to deepen their understanding of deep learning, from beginners to advanced practitioners. It provides a wealth of resources that can be used for both learning and reference, ensuring that the reader can revisit topics and find practical implementations easily.

Overview: This repository contains open-source projects from the Google AI Language team, focusing on language-related research and development. Key features: - Language models - Natural language processing - Text generation - Machine translation Tech stack: Python Why it matters: These projects contribute to advancing the field of artificial intelligence and natural language processing, offering valuable insights and tools for developers and researchers alike.

### Overview: 智能问答系统是人工智能和自然语言处理领域的重要研究方向，旨在通过快速、准确地获取信息来满足用户需求。北航大数据高精尖中心研究团队对基于知识图谱、文本、表格和视觉的问答系统进行了深入研究，并整合了学术界和工业界的理论和技术，形成了八个部分的总结。 ### Key features: - 社区问答(Community Question Answering, CQA) - 学术界 - 社区问答(Community Question Answering, CQA) - 工业界 - 机器阅读理解(Machine Reading Comprehension, MRC) - 学术界 - 机器阅读理解(Machine Reading Comprehension, MRC) - 工业界 - 知识图谱问答(Knowledge Base Question Answering, KBQA) - 学术界 - 知识图谱问答(Knowledge Base Question Answering, KBQA) - 工业界 - 表格问答(Table Question Answering, TQA) - 学术界 - 视觉问答(Visual Question Answering, VQA) - 学术界 ### Tech stack: Python, PyTorch, TensorFlow, NLTK, SpaCy, WordNet, Stanford CoreNLP, Spacy, Transformers, PyTorch Lightning, Hugging Face, Fairseq, FastText, GloVe, BERT, RoBERTa, DistilBERT, XLNet, Sentence-BERT, BERTweet, BERTopic, B

### Overview: AutoSwarm-Shadow-Adapter is a zero-modification Human-in-the-Loop adapter designed for OpenOPC's agentic DAG runtime. It allows AutoSwarm agents to pause non-blockingly when a role requires human sign-off or remote GPU execution, ensuring smooth operation without engine timeouts or parallel branch stalling. ### Key Features: - **Non-Blocking Interception**: Intercepts AutoSwarm task execution in milliseconds, parks the task in an isolated SQLite WAL database, and releases execution threads immediately with `AWAITING_HUMAN`. - **Contractor & Silicon Worker Portal**: Provides a dedicated, secure React UI and REST API for human reviewers and remote workers. - **Consolidated Corporate Brain**: Packages ancestor deliverables, structured markdown briefs, and uploaded file artifacts for downstream context inheritance. - **Canonical State Subordination**: Resumes AutoSwarm's native state machine immediately when deliverables are submitted. - **Symbiotic Architecture**: Functions as an independent extension that integrates with AutoSwarm's external agent registry with zero core modifications. ### Tech Stack: - Python 3.10+ - FastAPI - React 19 - SQLite WAL - JWT Auth ### Why It Matters: - **Non-Blocking Execution**: Ensures smooth AutoSwarm operations without engine timeouts. - **Secure and Reliable**: Provides a dedicated portal for human reviewers and remote workers. - **Context Inheritance**: Packages deliverables for downstream context inheritance. - **Simplified Integration**: Functions as an independent extension with zero core modifications, making it easy to integrate

### Overview: Ai2Thor is an open-source platform designed for visual AI research, focusing on embodied common sense reasoning and interaction. It provides a high-quality, interactive environment with 200+ custom-built scenes, 2600+ household objects, and multiple agent types including a custom LoCoBot, Kinova 3, and drone agents. The platform supports a wide range of actions and image modalities, making it ideal for studying embodied AI in various scenarios. ### Key Features: - **Custom Scenes:** 200+ high-quality scenes for exploration. - **Robust Objects:** 2600+ household objects with detailed annotations. - **Multi-Agent Support:** Multi-agent and custom agent types. - **Action Library:** 200+ actions for various interaction tasks. - **Image Modalities:** Support for RGB, instance segmentation, semantic segmentation, depth, normals, top-down, orthographic, and third-person camera frames. - **Scene Randomization:** Domain randomization within each scene. ### Tech Stack: C# ### Why It Matters: Ai2Thor is crucial for advancing research in embodied AI, providing a comprehensive and interactive environment for studying common sense reasoning, interaction, and embodied AI tasks. Its detailed scene and object annotations, along with the wide range of actions and image modalities, make it an essential tool for researchers in the field.

Overview: RAVE is an official implementation of the RAVE model, a real-time audio variational autoencoder designed for fast and high-quality neural audio synthesis. This project aims to provide a comprehensive and up-to-date resource for researchers and developers interested in the field of neural audio synthesis. Key features: - A detailed implementation of the RAVE model - Support for both regular and lazy dataset preparation methods - Training options for different configurations, including discrete and causal modes - Data augmentation capabilities for improved generalization in low-data regimes Tech stack: Python Why it matters: RAVE offers a practical and efficient solution for real-time audio synthesis, making it valuable for applications in music performance, installation art, and other fields where real-time audio synthesis is crucial. Repository: acids-ircam/RAVE GitHub stars: 1817 GitHub description: Official implementation of the RAVE model: a Realtime Audio Variational autoEncoder

Overview: This repository contains a survey paper on Text Classification Algorithms, focusing on text cleaning, pre-processing, and tokenization techniques. Key features: - Text cleaning and pre-processing methods - Tokenization of text documents - Example code for tokenization using NLTK Tech stack: Python Why it matters: This survey paper provides a comprehensive overview of text classification algorithms, emphasizing the importance of text cleaning and pre-processing in improving classification performance.

### Overview: This repository contains a comprehensive implementation of various semantic segmentation models using PyTorch. It includes models such as Deeplab V3+, GCN, UperNet, DUC, HDC, PSPNet, ENet, and FCN, along with a variety of datasets and loss functions. The project is designed to be highly configurable, allowing for easy tuning of parameters and support for different datasets and augmentations. ### Key Features: - **Models:** - **Deeplab V3+** - **GCN** - **UperNet** - **DUC** - **HDC** - **PSPNet** - **ENet** - **FCN** - **Datasets:** - **Pascal VOC** - **Semantic Contours from Inverse Detectors** - **Losses:** - Various loss functions for different models - **Learning Rate Schedulers:** - Support for different learning rate schedules - **Data Augmentation:** - Augmentation techniques to improve model performance ### Tech Stack: The project is implemented using PyTorch and Torchvision, along with additional libraries such as PIL for image processing and OpenCV for data augmentation. ### Why It Matters: Semantic segmentation models are crucial in computer vision tasks, enabling systems to identify and classify objects within images. This project provides a robust and flexible framework for implementing and experimenting with various segmentation models, making it valuable for researchers and developers looking to improve their segmentation algorithms.

### Overview: Edit Mind is a local-first video knowledge base that indexes and searches videos using multi-modal analysis and semantic natural language search. It is designed for privacy and ease of use, running on any computer or server with Docker installed. ### Key Features: - **Video Indexing and Processing:** A background service monitors for new video files and queues them for AI-powered analysis. - **AI-Powered Video Analysis:** Extracts metadata such as face recognition, transcription, object & text detection, scene analysis, and more. - **Vector-Based Semantic Search:** Powerful natural language search capabilities on video content using ChromaDB. ### Tech Stack: - Monorepo: [pnpm workspaces] - Containerization: [Docker], [Docker Compose] - Web Service: [React Router V7], [TypeScript], [Vite] - Background Jobs Service: [Node.js], [Express.js], [BullMQ] - ML Service: [Python], [PyAV], [PyTorch], OpenAI Whisper, Google Gemini or Ollama - Vector Database: [ChromaDB] - Relational DB: [PostgreSQL] (via [Prisma ORM]) ### Why It Matters: - Search videos by spoken words, objects, faces, etc... - Runs fully locally, respecting privacy. - Works on any computer or server with Docker installed. - Uses AI for rich metadata extraction and semantic search. ### Core Technologies: - **Monorepo:** [pnpm workspaces] - **Containerization:** [Docker], [Docker Compose]

### Overview: Agent Desktop provides a reliable way for agents to interact with desktop applications on macOS and Windows through the operating system's accessibility tree. It ensures that actions are performed safely and reliably, without relying on pixel-based guesses. ### Key Features: - **Command-Line Interface (CLI):** Agent Desktop is a command-line tool and C library that allows agents to interact with desktop applications. - **Snapshot Management:** Each element the agent can act on is assigned a unique reference (`@s8f3k2p9:e3`), which names the snapshot it came from. - **Stable Actions:** The actions remain stable and safe to retry, ensuring reliability and safety. - **Linux Support:** Plans to support Linux in the future. - **Command Set:** Ships with 60 command names, 56 operational commands, and the other four commands (`key-down`, `key-up`, `mouse-down`, `mouse-up`) are designed to work only with a daemon that can release held keys or buttons. - **Installation:** Can be installed globally using `npm install -g agent-desktop` or `pnpm add -g --allow-scripts=agent-desktop agent-desktop` on macOS and Windows. - **Permissions:** Requires specific permissions on macOS and Windows to interact with the operating system's accessibility tree and notifications. ### Tech Stack: Rust ### Why It Matters: Agent Desktop simplifies interaction with desktop applications, ensuring reliability and safety in agent-controlled environments.

Overview: BM25S is a fast implementation of BM25 in pure Python, optimized for sparse matrices and sparse scoring. It is designed to be both fast and simple to use, making it suitable for a wide range of applications, including text retrieval and search services. Key features: - Fast scoring: `bm25s` is optimized to handle large datasets efficiently, providing orders of magnitude performance improvement over other Python implementations. - Sparse matrix support: It uses sparse matrices to store eagerly computed scores, which speeds up the scoring process. - Easy to use: The library is designed to be simple and straightforward, making it accessible to developers without extensive knowledge of machine learning or NLP. Tech stack: `bm25s` is built using pure Python, with the use of Numpy for numerical computations and Numba for JIT compilation for speedup. Why it matters: `bm25s` is particularly useful in applications where performance is critical, such as in real-time search engines, text retrieval systems, or any scenario where fast and efficient ranking of documents based on a query is required. Its ability to handle large datasets and provide high throughput makes it an excellent choice for modern search and retrieval tasks. Repository: xhluca/bm25s Primary language: Python GitHub stars: 1814 GitHub description: Fast BM25 search in Python, powered by Numpy and Numba README excerpt: <div align="center"> <h1>BM25S⚡</h1> <i>BM25S (or BM2

### Overview: This roadmap aims to provide a comprehensive guide for beginners and aspiring Visual-SLAM developers, covering essential topics and skills needed to become proficient in the field. ### Key Features: - **Mathematics**: Basic probability, linear algebra, logarithms, calculus, and projective geometry. - **Programming**: C++ and Python, Bash/Linux. - **Camera/Imaging Fundamentals**: Pinhole camera model, camera calibration. - **Geometry and Optimization**: SLAM system anatomy, geometry, and optimization techniques. - **Monocular SLAM**: Classical methods like feature-based, direct, semi-direct, and structure from motion. - **RGB-D SLAM**: Dense tracking and volumetric/surfel fusion with depth sensors. - **Deep Learning + SLAM**: Frontend learning, differentiable backends, end-to-end systems, and neural SLAM. - **VIO/VINS**: Filtering vs optimization techniques. - **Stereo SLAM**: Metric scale and depth from stereo pairs. - **Collaborative SLAM**: Multi-robot mapping, loop closure, and map merging. - **LiDAR & Visual-LiDAR Fusion**: Odometry and tight camera-LiDAR-IMU fusion. - **Event Camera SLAM**: Asynchronous vision for HDR and high-speed motion. - **World Models & Spatial AI**: From SLAM maps to learned world representations. ### Tech Stack: Astro ### Why it Matters: This roadmap serves as a stepping stone for beginners, offering a structured approach to learning Visual-SLAM, from foundational concepts to advanced