Projects

### Overview: Join the RAGHack hackathon, where you can build AI applications using RAG technology. The hackathon features 25+ live streams, covering Python, Java, JavaScript, C#, and more, with Azure AI Studio, LangChain, Semantic Kernel, and multiple retrievers. The hackathon starts September 3rd and ends September 13th, with hack submissions due September 16th, 11:59 PM PST. ### Key Features: - Build RAG applications using Python, Java, JavaScript, C#, and more. - Explore Azure AI Studio, LangChain, Semantic Kernel, and multiple retrievers. - Learn about the latest technology, including agents and vision models. - Submit your hack for a chance to win prizes. ### Tech Stack: - Python, Java, JavaScript, C#, and more. - Azure AI Studio, LangChain, Semantic Kernel, and multiple retrievers. ### Why It Matters: The possibilities are endless when it comes to what you can create with RAG technology. The hackathon offers a chance to learn from experts and build real-world applications. Join the hackathon and start building your RAG application today!

### Overview: 深度强化学习（Deep Reinforcement Learning）与“羊了个羊”游戏结合，探索AI技术与休闲游戏的创新可能。 ### Key features: - 强化学习算法训练 - 本地部署和在线试玩功能 - 神经网络模型和Actor-Critic架构 - 3tiles游戏环境和策略设计 ### Tech stack: Python ### Why it matters: 通过结合深度强化学习和休闲游戏，探索AI在娱乐领域的创新应用，提高用户体验和游戏的趣味性。 Repository: opendilab/DI-sheep GitHub stars: 522 GitHub description: 羊了个羊 + 深度强化学习（Deep Reinforcement Learning + 3 Tiles Game) README excerpt: # DI-sheep：深度强化学习 + 羊了个羊 当 AI 技术的明珠——深度强化学习，遇到号称“通关率只有0.01%”的游戏“羊了个羊”，会碰撞出哪些奇思妙想呢？ P.S. 路过记得点个 star  ，持续更新ing。 P.S.S. 想了解更多深度强化学习相关知识？快来 [DI-engine](https://github.com/opendilab/DI-engine) 训练自己的智能体。 <div align="center"> <a href="https://github.com/op

### Overview: Score Jacobian Chaining: Lifting Pretrained 2D Diffusion Models for 3D Generation is a project that leverages the capabilities of pretrained 2D diffusion models to generate 3D data. This innovative approach aggregates 2D scores from multiple camera viewpoints into a 3D score, utilizing a voxel radiance field to repurpose a pretrained 2D model for 3D data generation. The project addresses a technical challenge related to distribution mismatch and proposes a novel estimation mechanism to resolve it. ### Key Features: - **Score Aggregation**: Combines 2D scores from multiple camera viewpoints into a 3D score. - **Pretrained 2D Model Repurposing**: Utilizes a pretrained 2D model to generate 3D data. - **Jacobian Chain Rule Application**: Applies the chain rule on the learned gradients to back-propagate the score through the Jacobian of a differentiable renderer. - **Technical Challenge Resolution**: Identifies and resolves a challenge related to distribution mismatch in the application. ### Tech Stack: - Python ### Why It Matters: This project demonstrates the potential of combining pretrained models for generative tasks, showcasing how 2D models can be repurposed for 3D data generation. It addresses a specific technical challenge related to distribution mismatch and proposes a solution that could be applicable to other similar applications.

Overview: LEDNet is a lightweight encoder-decoder network designed for real-time semantic segmentation, achieving state-of-the-art results on the Cityscapes dataset with less than 1M parameters and over 71 FPS on a single GTX 1080Ti GPU card. Key features: - Lightweight architecture with less than 1M parameters - Asymmetric encoder-decoder design - Channel split and shuffle operations in the encoder to reduce computational cost - Attention pyramid network (APN) in the decoder to further lighten the network complexity - Real-time semantic segmentation capability Tech stack: Python, PyTorch, CUDA, PyTorch-CUDNN Why it matters: LEDNet offers a practical solution for real-time semantic segmentation tasks, making it an effective method for applications requiring high-speed processing and accurate results.

### Overview: ARCTIC is a comprehensive dataset designed for studying Dexterous Bimanual Hand-Object Manipulation. This repository provides tools and resources for downloading, processing, visualizing, and training models on the ARCTIC dataset. The dataset includes 2.1M high-resolution images paired with annotated frames, enabling large-scale machine learning tasks. ### Key Features: - **Image Processing**: Tools for preprocessing, splitting, and visualizing images. - **Model Training**: Code to reproduce baseline models and develop custom models for the CVPR 2023 paper. - **Template-Free Reconstruction**: Methods for reconstructing hands and objects without using templates. - **Mixed-Reality Setup**: Images captured from 8x 3rd-person views and 1x egocentric view. - **MoCap Data**: High-quality 13k registered 3DMD hand scans and 90k calibrated multiview RGB images. - **3D Groundtruth**: 3D groundtruth for SMPL-X, MANO, and articulated objects. - **High-Dexterous Motion**: Contains highly dexterous bimanual manipulation motion beyond quasi-static grasping. ### Tech Stack: Python ### Why it Matters: ARCTIC is a crucial resource for researchers and practitioners interested in Dexterous Bimanual Hand-Object Manipulation. It provides a rich dataset for training models and benchmarking performance, making it an essential tool in the field of computer vision and robotics. ### Repository: Repository: zc-alexfan/arctic GitHub

### Overview: TensorFlow 2.x offers a more user-friendly and efficient experience with features like eager execution, unified APIs, and seamless integration with Python. It simplifies model building and training, making it easier to get started with deep learning projects. ### Key Features: - **Eager Execution**: Runs computations directly in the Python runtime, eliminating the need for eager execution. - **Unified APIs**: Simplifies model building with unified RNNs and optimizers, making it easier to switch between different models. - **Flexibility**: Supports models on any platform, including CPUs and GPUs. ### Tech Stack: Python ### Why It Matters: TensorFlow 2.x is designed to be more intuitive and easier to use, making it a preferred choice for developers looking to quickly get started with deep learning projects. Its flexibility and ease of use make it a valuable resource for both beginners and experienced practitioners in the field.

Overview: GenSX is a TypeScript framework designed for building complex LLM applications, including agents, chatbots, and long-running workflows. It simplifies the process of creating workflows by composing components and provides features like type safety, automatic tracing, and built-in storage. Key features: - Pure functions for components - Natural composition of workflows - Type-safe and built-in support for LLMs - Built-in deployment and storage - Automatic tracing for debugging and observability Tech stack: TypeScript Why it matters: GenSX simplifies the development of complex LLM applications by providing a straightforward workflow engine that is easy to learn and extend. It offers a clean, type-safe approach to building workflows, making it suitable for both beginners and experienced developers. Repository: gensx-inc/gensx GitHub stars: 522 GitHub description: The TypeScript framework for agents & workflows with react-like components. Lightning fast dev loop. Easy to learn. Easy to extend.

Overview: Phonetisaurus G2P is a repository containing scripts and tools for training, evaluating, and using grapheme-to-phoneme models for speech recognition using the OpenFst framework. Key features: - C++ binaries for training, compiling, and evaluating G2P models - Simple Python bindings for extracting individual multigram scores, alignments, and dumping raw lattices in .fst format - Scripts and bindings tested with Python 3.8.5 - Standalone distributions for previous INTERSPEECH papers and the complete, exported final version of the old Google Code repository Tech stack: Shell Why it matters: Phonetisaurus G2P provides a comprehensive set of tools for speech recognition, making it an essential resource for researchers and developers working in the field of speech processing and machine learning.

### Overview: Teamlinker is a comprehensive team collaboration platform that integrates multiple functional modules, including project management, wiki, calendar, meeting, chat, and file sharing, to enhance team collaboration efficiency and streamline project management. ### Key features: - **Project Management**: Customizable work item types, workflows, fields, Kanban, and Gantt chart management. - **Wiki**: Supports rich text editing, color and font adjustments, link and picture insertion, and attachment management. - **Calendar**: Multi-calendar management, time zone switching, day, week, and month view, all-day and repeated event support, and calendar event reminders. - **Meeting**: Real-time video and voice transmission, member invitation, screen sharing, virtual background, and chat functionality. - **Chat**: Fast video conferencing and message collection/search. - **File Management**: Seamless integration with network disk, allowing users to manage files and collaborate on projects. ### Tech stack: TypeScript ### Why it matters: Teamlinker provides a unified platform for efficient team collaboration, reducing the need for multiple tools and improving project management efficiency. The platform is built on a web operating system (TeamOS), allowing for parallel task processing and seamless integration, making it a cost-effective solution for enterprise users. Repository: Teamlinker/Teamlinker GitHub stars: 521 GitHub description: Teamlinker is a team collaboration platform that integrates multi-functional modules. Users can process tasks in parallel, including six functional modules: project, wiki, calendar, meeting, chat and network disk, achieving seamless integration and improving team collaboration

Overview: This project is a Keras implementation of the Deep Semantic Similarity Model (DSSM) and Convolutional Latent Semantic Model (CLSM), as described in the research paper by Microsoft Research. Key features: - Keras implementation - Deep semantic similarity model - Convolutional latent semantic model Tech stack: Python Why it matters: This project provides a practical tool for implementing and utilizing the DSSM and CLSM models, which are useful for information retrieval and semantic similarity analysis.

### Overview: The repository contains a comprehensive list of supervised and unsupervised depth completion methods, focusing on the VOID and KITTI benchmarks. It provides an overview of recent research in depth completion, including metrics such as MAE, RMSE, iMAE, and iRMSE. ### Key Features: - **Unsupervised VOID Benchmark**: Compares methods like AugUndo, Monitored Distillation, DesNet, and Unsupervised Depth Completion with Calibrated Backprojection Layers. - **Supervised VOID Benchmark**: Includes methods such as KBNet and ScaffNet. - **Unsupervised KITTI Benchmark**: Focuses on methods like ScaffNet and DesNet. - **Supervised KITTI Benchmark**: Covers methods like KBNet and ScaffNet. ### Tech Stack: The repository uses PyTorch and TensorFlow for implementing the methods. ### Why It Matters: The repository serves as a valuable resource for developers interested in improving depth completion techniques, providing insights into state-of-the-art methods and their performance metrics.

Overview: OME is a Kubernetes operator designed for enterprise-grade management and serving of Large Language Models (LLMs). It optimizes deployment and operation through model management, intelligent runtime selection, resource optimization, and advanced deployment patterns. Key features: - Model Management: Custom resources and automated model parsing. - Intelligent Runtime Selection: Weighted scoring for optimal model deployment. - Optimized Deployments: Multiple deployment patterns with advanced scaling controls. - Resource Optimization: Specialized GPU scheduling and dynamic re-optimization. - Runtime Integrations: Support for SGLang and vLLM. - Accelerator Management: Hardware-aware scheduling and intelligent accelerator selection. - Fleet Capacity and GPU Operations (alpha): Optional add-on control planes. - Command-Line Interface: Official CLI for inspecting and managing OME resources. - Web Console: Modern web interface for monitoring and managing OME. Tech stack: Go Why it matters: OME simplifies the deployment and management of LLMs, optimizing resource usage and providing advanced features for enterprise environments. Repository: ome-projects/ome GitHub stars: 517

### Overview: Deuz SDK is a TypeScript-based AI SDK and agent framework that provides zero dependencies, enabling the creation of durable, multi-agent swarms with features like long-term memory, hybrid RAG, MCP, and more. It supports various providers such as Claude, GPT, Gemini, Grok, DeepSeek, and Ollama. ### Key features: - **Models**: Multiple providers, including Claude, GPT, Gemini, Grok, DeepSeek, and Ollama, with streaming and structured output. - **Agents**: Typed results, verification, approvals, resumable runs, and memory management. - **Swarms**: Task dependencies, runtime spawning, shared blackboards, rounds, and persistence with SQLite, Postgres, or Redis. - **Operations**: Leases, cross-process cancel and recovery, persistent budgets, schedules, goal records with a goal supervisor, and a local Docker runner (preview). - **Evolve**: Evolutionary program search with a mandatory budget and zero-call resume. - **Context**: Versioned memory, context compiler, retrieval, compaction, transcript search, and MCP tools. - **Interop**: MCP server, A2A to call other agents, and serve your own. - **Control**: Shared execution policies, budget accounting, traces, eval suites, and optional React bindings. ### Tech stack: TypeScript ### Why it matters: Deuz SDK simplifies the creation of AI agents and multi-agent swarms, providing a robust and scalable solution for developers without the need for external dependencies.

Overview: Goopt is an experimental project that simulates a procedural web using GPT-3, focusing on generating synthetic content for the web. Key features: - Procedural web simulation - GPT-3-based content generation - Search engine functionality Tech stack: JavaScript Why it matters: Goopt explores the potential of the procedural web, offering infinite content generated by algorithms and AI, challenging traditional web creation and enhancing cybernetic capabilities. Repository: jokenox/Goopt Primary language: JavaScript GitHub stars: 520 GitHub description: 🔍 Search Engine for a **Procedural Simulation** of the **Web** with **GPT-3**. README excerpt: <div align="center">    **Search Engine** for a **Procedural Simulation** of the **Web** with **GPT-3** </div> --- > **Web 4.0** could be the propitious evolution for the **procedural web**. Goopt is an experiment in what the "**procedural web**" could be. This new web will use **procedural content generation** to create varied content, **completely synthetic**, since these are **generated by algorithms** and **artificial intelligence**. For now, the content is only text that is being

Overview: This repository is a curated list of papers and resources related to Retrieval-Augmented Generation (RAG) for large language models (LLMs). It focuses on the development and advancements in the field. Key features: - Papers on RAG for LLMs - Papers on NIPS, EMNLP, ACL, ICML, and ICLR conferences - Papers on recent developments in RAG for LLMs - Papers on 2025, 2026, 2024, and 2025 papers Tech stack: Python, Markdown, and GitHub Why it matters: This repository serves as a comprehensive resource for researchers and developers interested in the latest advancements in RAG for LLMs, providing a platform to stay updated with the latest research and developments in the field.

### Overview: This repository contains a collection of useful datasets for robotics and computer vision research. The datasets are categorized into robotics, computer vision, and others, providing a comprehensive resource for researchers and practitioners in the field. ### Key Features: - **Robotics Datasets:** - **Malaga Datasets**: Included in the MRPT repository. - **IJRR Data Papers**: Contains classic datasets from the International Journal of Robotics Research. - **Awesome SLAM Datasets**: Managed by Younggun Cho. - **YACVID**: Yet Another Computer Vision Index To Datasets. - **Computer Vision Datasets on the Web**: Various datasets from CVonline and CVPapers. - **CVonline Image Databases**: Datasets from CVonline. - **Computer Vision Online Datasets**: Datasets from Computer Vision Online. - **Machine Learning Repository**: Datasets from UCI. - **Kaggle Datasets**: Datasets from Kaggle. - **IEEE DataPort**: Datasets from IEEE. - **Computer Vision Datasets:** - **CVonline Image Databases**: Datasets from CVonline. - **Computer Vision Datasets on the Web**: Various datasets from CVonline and CVPapers. - **YACVID**: Yet Another Computer Vision Index To Datasets. - **Computer Vision Online Datasets**: Datasets from Computer Vision Online. - **CVonline Image Databases**: Datasets from CVonline. - **Computer Vision Datasets on the Web**: Various datasets from CVonline and

### Overview: This survey explores **Vision-Language-Action Models (VLA)** for autonomous driving, focusing on models that integrate vision, language, and action processing. It traces the evolution from early Vision-Action (VA) models to modern VLA frameworks, categorizing them into End-to-End VLA and Dual-System VLA paradigms. ### Key Features: - **End-to-End VLA**: Models that integrate perception, reasoning, and planning within a single framework. - **Dual-System VLA**: Models that separate slow deliberation (via Value Learning Machines, VLMs) from fast, safety-critical execution (via planners). ### Tech Stack: - **End-to-End VLA**: Utilizes a unified architecture for perception, reasoning, and planning. - **Dual-System VLA**: Separates slow deliberation and fast execution, enhancing performance and safety. ### Why It Matters: - **Improved Performance**: Combines vision, language, and action processing to enhance decision-making capabilities. - **Safety-Critical Execution**: Separates slow deliberation from fast execution to ensure safety and efficiency. Repository: worldbench/awesome-vla-for-ad Primary Language: HTML GitHub Stars: 518 GitHub Description: 🌐 Vision-Language-Action Models for Autonomous Driving: Past, Present, and Future README Excerpt: [![Awesome Logo](https://cdn.rawgit.com/sindresorhus/awesome/d7305f38d29fed78fa85652e

Overview: This project is a Mac-compatible offline voice assistant that utilizes Mistral 7b via Ollama and Whisper speech recognition models. Key features: - Mac compatibility - High-quality voice assistant - Offline functionality Tech stack: Python Why it matters: This project offers a practical and efficient solution for voice assistance on macOS, enhancing accessibility and convenience for users. Repository: apeatling/ollama-voice-mac Primary language: Python GitHub stars: 519

Overview: MAV, or Model Activity Visualiser, is a tool designed to help developers visualize the internal workings of large language models, specifically focusing on how they generate text. This visualization is crucial for understanding the model's behavior and improving model performance. Key features: - Real-time model visualization - Support for various Hugging Face models including GPT-2, SmolLM-135M, and others - Customizable panels for different types of data visualization - Interactive mode for enhanced user experience - Ability to run with a training loop and custom model configurations Tech stack: MAV uses Python as the primary programming language, leveraging the PyTorch framework for model training and visualization. Why it matters: MAV provides a valuable tool for developers to better understand and optimize the performance of large language models, making them more effective and efficient in their applications.

Overview: Embedbase is a simple API designed to integrate LLMs and VectorDBs, making it easy to build LLM-powered applications without the need for hosting. Key features: - Generate: Use `.generateText()` to leverage 9+ LLMs. - Semantic Search: Use `.add()` to create semantically searchable information and `.search()` to run semantic queries. Tech stack: TypeScript Why it matters: Embedbase simplifies the integration of LLMs and VectorDBs, allowing developers to build applications quickly and efficiently without the complexities of hosting and managing their own infrastructure.