Projects

Overview: HyperQwen is a fast and efficient solution for serving large Qwen models on GPUs that people actually own, using vLLM. Key features: - Supports Qwen3.8-27B model on a single 24 GB consumer GPU with vLLM. - Available in single-user mode with 150k token context and OpenAI-compatible API. - Supports batch mode with 1,035 tok/s aggregate decode at 64 concurrent requests. - Provides a patch series against a pinned vLLM and a model-preparation pipeline. Tech stack: Python. Why it matters: HyperQwen is designed to maximize the performance and efficiency of Qwen models on GPUs, making it a valuable tool for developers and researchers looking to deploy large models on consumer hardware. Repository: syv-ai/HyperQwen Primary language: Python GitHub stars: 2070 GitHub description: Serve large Qwen models fast on the GPUs people actually own. Qwen3.8-27B on a single 24 GB card with vLLM: 127 tok/s single-user (381 when the answer quotes the prompt), ~1,035 tok/s at 64 concurrent, 150k-262k context. vLLM patches, requant pipeline, benchmarks. README excerpt: <div align="center"> <h1>HyperQwen</h1> <p><b>Large Qwen models, served fast on the GPUs people actually own.</b

### Overview: This repository is an initiative to compile and catalog resources that utilize machine learning techniques to address combinatorial optimization problems. Contributions are marked by the Thinklab team, and the project aims to provide a comprehensive overview of various algorithms and methods used in this field. ### Key Features: - **Survey Papers**: Comprehensive reviews of machine learning approaches to combinatorial optimization. - **Problems**: Detailed breakdowns of common combinatorial optimization problems such as the Travelling Salesman Problem (TSP), Job Shop Scheduling Problem (JSSP), Flow Shop Problem (FSP), Sorting & Ranking (Sort&Rank), Graph Matching (GM), Quadratic Assignment Problem (QAP), Portfolio Optimization (PortOpt), Maximal Cut, Vehicle Routing Problem (VRP), Maximum Independent Set, Generalization, Orienteering Problem (OP), Knapsack, Boolean Satisfiability (SAT), Computing Resource Allocation, Bin Packing Problem (BPP), Graph Edit Distance (GED), Hamiltonian Cycle Problem (HCP), Graph Coloring, Maximal Common Subgraph (MCS), Influence Maximization, Max Clique, Mixed Integer Programming (MIP), Causal Discovery, Differentiable Optimization, Car Dispatch, Electronic Design Automation (EDA), Conjunctive Query Containment, Virtual Network Embedding (VNE), Predict+Optimize, Optimal Power Flow, Facility Location Problem (FLP), and Combinatorial Drug Recommendation. ### Tech Stack: Python ### Why It Matters: This repository serves as a valuable resource for researchers, practitioners, and students interested in machine learning applications

Overview: Nestia is a comprehensive set of libraries and tools designed to enhance the development of NestJS applications, including advanced features like WebSocket routes, Swagger generator, and automatic E2E test functions. Key features: - `@nestia/core`: Super-fast/easy decorators and advanced WebSocket routes. - `@nestia/sdk`: Swagger generator, SDK library generator, and mockup simulator for client applications. - `@nestia/e2e`: Test program utilizing E2E test functions. - `@nestia/benchmark`: Benchmark program using E2E test functions. - `@nestia/editor`: Swagger-UI with an online TypeScript editor. - `@agentica`: Agentic AI library for specialized LLM function calling. - `@autobe`: Vibe coding agent for generating NestJS applications. - `nestia`: CLI tool for command line interface. Tech stack: TypeScript. Why it matters: Nestia significantly enhances NestJS application development by providing optimized decorators, advanced routing, and powerful testing tools, making it a valuable addition to the developer's toolkit. Repository: samchon/nestia GitHub stars: 2176 GitHub description: NestJS Helper + AI Chatbot Development GitHub license: MIT Downloads: 2176

Overview: Intel® Extension for Transformers is a toolkit designed to accelerate GenAI/LLM models, offering state-of-the-art compression techniques and efficient inference capabilities on Intel platforms. Key features: - Support for Qwen2 models - Support for Meta Llama 3 - INT4 inference on Intel Meteor Lake - 1.8x performance improvement on GPT-J inference - Support for INT4 inference on Intel GPUs, including Intel Data Center GPU Max Series and Intel Arc A-Series - Support for 4-bit chatbot demo - Support for 7B-sized LLM NeuralChat-v3-1 and DPO dataset Tech stack: Python Why it matters: Intel® Extension for Transformers simplifies the development of AI/LLM models, making it easier to deploy high-performance models on Intel hardware, including CPUs and GPUs, across various devices. This toolkit leverages Intel's expertise in AI acceleration to achieve significant performance improvements, making it an essential tool for developers looking to leverage the power of Intel platforms for AI applications.

### Overview: This project provides a speech recognition system using TensorFlow, utilizing sequence-to-sequence neural networks. It replaces the previous implementation with Google's TensorFlow framework and sequence-to-sequence models. ### Key Features: - **Speech Recognition**: Implements speech recognition using TensorFlow deep learning framework and sequence-to-sequence neural networks. - **Compatibility**: Compatible with TensorFlow 1.0, but the project is no longer actively maintained. - **Model Training**: Includes models and training data for speech recognition, such as 100GB of synthetic Text-to-Speech snippets and 21GB of YouTube captions. ### Tech Stack: Python ### Why It Matters: This project serves as an educational resource for developers looking to implement speech recognition systems using TensorFlow. It provides a practical example of how to build a speech recognition model from scratch, which can be beneficial for understanding deep learning architectures and TensorFlow's capabilities. ### Repository: Repository: pannous/tensorflow-speech-recognition GitHub Stars: 2174 GitHub Description: 🎙 Speech recognition using the tensorflow deep learning framework, sequence-to-sequence neural networks.

### Overview: This repository is dedicated to resources for generating JSON using Large Language Models (LLMs) via function calling, tools, CFG, libraries, models, and notebooks. ### Key Features: - Function Calling - Local Models - Python Libraries - Blog Articles - Videos - Jupyter Notebooks - Leaderboards ### Tech Stack: LLMs, Function Calling, Local Models, Python Libraries, Blogs, Videos, Notebooks, Leaderboards ### Why it Matters: This list provides a comprehensive collection of resources and tools for leveraging LLMs to generate structured outputs, including JSON, XML, and YAML. It covers various models, libraries, and tools, making it a valuable resource for developers and researchers in the field of natural language processing and AI.

### Overview: WorldGen is a powerful tool for generating 3D scenes in seconds from text prompts and images. It supports instant 3D generation, 360° free exploration, diverse scene support, and flexible rendering. Ideal for games, simulations, robotics, and virtual reality applications. ### Key features: - **Instant 3D Generation** ⚡️: Create full 3D scenes from input data in seconds. - **360° Free Exploration** 🧭: WorldGen supports free 360° consistent exploration of the generated 3D scene with loop closure. - **Diverse Scenes Support** 🌈: WorldGen supports both indoor and outdoor scenes, realistic and unrealistic scenes in any style. - **Flexible Rendering** 📸: WorldGen supports rendering at any resolution with any camera setting and trajectory in real-time. ### Tech stack: Python ### Why it matters: WorldGen is a valuable tool for anyone looking to create 3D environments and scenes quickly and efficiently. Whether you're a game developer, a simulation expert, or a researcher in robotics, WorldGen can help you generate high-quality 3D scenes in seconds, making your projects faster and more efficient. ### Repository: - **Primary Language**: Python - **GitHub Stars**: 2,170 - **Repository**: ZiYang-xie/WorldGen - **Author**: Ziyang Xie - **Contact Email**: ziyangxie01@gmail.com - **News and TODOs**: Updated to support ml

Overview: SpikingJelly is an open-source deep learning framework specifically designed for Spiking Neural Network (SNN) applications, built on PyTorch. It supports large-scale SNN training and inference, offering features like operator execution, event-based datasets, and automatic execution. Key features: - Spiking neurons - Surrogate gradients - Stateful and stateless modules - PyTorch-native training flows - ANN2SNN conversion - Automatic execution with Torch, Triton, and native CUDA - Memory-efficient training - Distributed training - Neuromorphic datasets Tech stack: Python Why it matters: SpikingJelly simplifies the development of SNN applications, making it easier to create and train SNNs using PyTorch, while also providing a comprehensive set of tools for efficient and scalable SNN training and inference.

Overview: MRPT is a robust C++ framework for mobile robotics, developed since 2005. It has been widely used in research and industry. Key features: - SLAM: MonteCarlo Localization, RBPF, pose-graph optimization - Geometry: SE(2)/SE(3) Lie groups, rigid body transforms, point clouds - Probabilistic: Kalman filters, particle filters, PDFs over poses & maps - Maps: Occupancy grids, point maps, landmark maps, voxel maps - Sensors: LiDAR, stereo/RGB-D cameras, IMU, GPS, Velodyne, ... - Visualization: 3-D scene graph, GUI windows, plot widgets - Navigation: Reactive nav, path planning, PTG-based obstacle avoidance - Python: pybind11 bindings for all major modules Tech stack: C++ Why it matters: MRPT is a well-established toolkit for mobile robotics, with over 2,169 GitHub stars, used in both research and industry. It provides a robust and modular framework for various robotics tasks, including SLAM, geometry, probabilistic methods, and sensor integration.

Overview: AI Chatbot Framework is a self-hosted, low-code platform for building conversational AI chatbots using Python, designed for developers without extensive AI expertise. Key features: - Fully self-hosted - Low-code admin dashboard for bot development - Multi-turn conversations - API request fulfillment - Persistent memory and context management - Advanced Natural Language Understanding (NLU) with Spacy, ML-based intent recognition, entity extraction, and zero-shot NLU using large language models - Knowledge base and FAQ answering using RAG (in development) - Conversation logs - Channel integrations: Web via REST API/Chat Snippet, Facebook Messenger, Slack (coming soon), WhatsApp via Twilio (coming soon) Tech stack: Python / FastAPI / Pydantic / MongoDB / Motor / React / NextJS / scikit-learn / Tensorflow / Keras / Spacy / python-crfsuite / Docker / docker-compose / Kubernetes / Helm Why it matters: AI Chatbot Framework simplifies the creation of AI-powered chatbots, enabling developers to harness the power of AI without needing advanced coding skills, making it accessible for a wide range of applications and industries.

### Overview: JAX is a library that brings automatic differentiation and the XLA compiler together through a NumPy-like API, enabling high-performance machine learning research on GPUs and TPUs. ### Key features: - **Automatic Differentiation**: JAX automatically computes gradients and derivatives, making it easy to optimize machine learning models. - **XLA Compiler**: JAX integrates the XLA compiler, which optimizes machine learning models for faster execution on accelerators like GPUs and TPUs. - **NumPy-like API**: JAX provides a familiar NumPy-like interface for building and training machine learning models, making it accessible to researchers and practitioners. ### Tech stack: JAX is built on top of TensorFlow, NumPy, and XLA, leveraging the strengths of each to provide a powerful and flexible framework for machine learning research. ### Why it matters: JAX is crucial for researchers and practitioners who need to run machine learning models on GPUs and TPUs, providing a high-performance, efficient, and scalable solution.

### Overview: Oh My PPT 是一款由 TypeScript 语言驱动的 AI PPT、配图生成与编辑工具。它通过本地优先的工作流，帮助用户从创作到编辑、演示与导出整个过程都在一套本地系统中完成，提供灵活的模型选择和透明的中间层支持。 ### Key features: - **本地优先的工作流**：从创作到编辑、演示与导出，都在一套本地系统中完成。 - **AI驱动的可编辑 HTML**：通过 AI 驱动的 HTML 生成，重构下一代 PPT 生产方式。 - **描述你想表达的内容**：让 AI 完成大纲、页面与配图。 - **从创作到编辑、演示与导出**：在一套本地系统中完成整个过程。 ### Tech stack: TypeScript ### Why it matters: 通过本地优先的工作流，Oh My PPT 提供了一个灵活且透明的解决方案，帮助用户高效地创建、编辑和展示 PPT，同时保持模型的原生输出体验。

Overview: Fugue is a unified interface for distributed computing that allows users to execute Python, Pandas, and SQL code on Spark, Dask, and Ray with minimal rewrites. Key features: - Bringing Python and Pandas code to Spark, Dask, and Ray without rewrites - Using FugueSQL to define end-to-end workflows on top of Pandas, Spark, and Dask DataFrames - Minimal rewrites for parallelizing or scaling existing Python and Pandas code Tech stack: Python Why it matters: Fugue simplifies distributed computing by providing a unified interface for executing Python, Pandas, and SQL code on Spark, Dask, and Ray, making it easier to parallelize and scale existing code without the need for extensive rewrites.

### Overview: This project is an experimental implementation of synthesizing images from captions using Skip Thought Vectors, utilizing the GAN-CLS Algorithm. The images are generated using the DCGAN in TensorFlow, with support for datasets like the flowers dataset. The model architecture is detailed in the README, and the project is built on Python 2.7.6, TensorFlow, and other libraries. ### Key Features: - **Text to Image Synthesis**: Converts textual descriptions into images. - **Skip Thought Vectors**: Utilizes a neural network to capture the meaning of captions. - **GAN-CLS Algorithm**: Combines Generative Adversarial Networks with Conditional Label Smoothing for improved image generation. - **Datasets**: Supports the flowers dataset for training and testing. - **Model Architecture**: Detailed in the README for easy understanding. - **Requirements**: Python 2.7.6, TensorFlow, h5py, Theano, scikit-learn, and NLTK. - **Usage**: Data processing, training, and generating images from captions. ### Tech Stack: - Python 2.7.6 - TensorFlow - h5py - Theano - scikit-learn - NLTK ### Why It Matters: This project offers a novel approach to image generation based on textual descriptions, potentially opening up new avenues in image synthesis and AI research. The Skip Thought Vectors technique adds an interesting layer of semantic understanding to the image generation process, making it more contextually relevant. ### Repository: - **GitHub Stars**: 2166 -

### Overview: React Native Voice Recognition library for iOS and Android (Online and Offline Support) ### Key Features: - **Speech Recognition**: Transcribes spoken words into text. - **Offline Support**: Works offline, even without internet connectivity. - **Cross-Platform**: Supports both iOS and Android. - **Modern Speech Recognition**: Utilizes modern speech-to-text technology. ### Tech Stack: TypeScript ### Why it Matters: This library simplifies the process of integrating speech recognition into your React Native application, making it easier to build voice-based features without the need for complex native code. It supports both online and offline modes, ensuring your app remains functional in various environments.

### Overview: NannyML is an open-source Python library designed for post-deployment data science. It provides tools to estimate model performance, detect data drift, and link data drift alerts to changes in model performance. NannyML is particularly useful for data scientists looking to maintain visibility and trust in their deployed machine learning models. ### Key Features: - **Performance Estimation**: Estimates model performance without access to targets. - **Data Drift Detection**: Detects data drift using multivariate data reconstruction with PCA. - **Interactive Visualizations**: Provides easy-to-use interfaces and interactive visualizations for better understanding. ### Tech Stack: - Python ### Why It Matters: NannyML empowers data scientists to quickly identify and address silent model failures, ensuring continuous model performance monitoring and trust in deployed models. By automating the process of detecting and analyzing data drift, NannyML helps prevent model degradation and ensures the reliability of deployed machine learning models. ### Repository: NannyML/nannyml - **Primary Language**: Python - **GitHub Stars**: 2160 ### GitHub Description: NannyML: Post-deployment data science in Python ### Documentation: - Website: [nannyml.com](https://nannyml.com/) - Docs: [nannyml.readthedocs.io](https://nannyml.readthedocs.io/en/stable/) - Community Slack: [join.slack.com/t/nannymlbeta/shared_invite/zt-16fvpeddz-HAvTsjNEyC9CE6JXbiM7BQ](https

### Overview: Image-to-LaTeX converts images of LaTeX math equations into LaTeX code, leveraging preprocessed datasets and modern transformer-based models to improve performance and generalization. ### Key Features: - **Model Architecture**: Uses ResNet-18 encoder with 2D positional encoding and a Transformer decoder for efficient image-to-markup generation. - **Data Preprocessing**: Raw dataset with image augmentation for increased diversity and adaptability. - **Normalization**: Normalizes LaTeX code to ensure visually identical outputs. - **Space Handling**: Removes space commands and blank images to prevent generating LaTeX code on empty images. ### Tech Stack: Python ### Why it Matters: This project addresses the challenge of converting images of LaTeX math equations into LaTeX code, addressing a common need in scientific and technical writing. By leveraging modern transformer models and improved data preprocessing, it offers better performance and generalization compared to previous methods.

### Overview: This guidebook provides practical insights and theoretical knowledge about language model evaluation, covering various evaluation methods, designing evaluations, and tips and tricks from real-world experience. ### Key features: - **Basics**: Introduction to evaluation methods and concepts. - **Designing your automatic evaluation**: Guidance on creating automated evaluation metrics. - **Some evaluation datasets**: Resources for using specific datasets in evaluations. - **Tips and tricks**: Practical advice and tips for improving evaluation results. - **Human evaluation**: How to use human annotators effectively and tips for improving human evaluations. - **LLM-as-a-judge**: How to design and evaluate your own evaluation prompt, and tips for evaluating your evaluator. - **Troubleshooting**: Solutions for common issues in inference and reproducibility. - **General knowledge**: Beginner-level guides to LLM basics, with tips and references for advanced users. ### Tech stack: Jupyter Notebook ### Why it matters: Whether you're working with production models, a researcher, or a hobbyist, this guidebook provides essential knowledge and practical tips to ensure your language models perform well on specific tasks.

Overview: Neurite is an open-source project designed to enhance digital interfaces by integrating fractal graph theory with contemporary mind-mapping techniques. It aims to create a limitless workspace that blends the complexity of fractals with modern mind mapping, focusing on non-integer dimensional thought. Key features: - FractalGPT integration - Multi-agent UI - Neurite Desktop - Neural API Tech stack: JavaScript Why it matters: Neurite leverages fractal logic to explore the hidden patterns and intricate connections in creative thinking, challenging conventional notions of four-dimensional space and fostering a unique approach to collaboration and knowledge management. Repository: satellitecomponent/Neurite Primary language: JavaScript GitHub stars: 2155 GitHub description: Fractal Graph-of-Thought. Rhizomatic Mind-Mapping for Ai-Agents, Web-Links, Notes, and Code. README excerpt: 🌐 **[neurite.network](https://neurite.network/)** 🌐 This is an open-source project in active development. The Mandelbrot Set is not just an aesthetic choice - fractal logic is ingrained into a countless number of natural and constructed phenomena, including polynomial equations, art, music, and the cosmic web. Fractals act as the cross-disciplinary framework for non-integer dimensional thought, challenging conventional notions of four-dimensional spacetime. Neurite aims to grow agentic graphs of fractal creativity & collaboration. 🧩 **Drawing from chaos theory and graph theory, Neurite unleashes the hidden patterns and

### Overview: ReAgent is an open-source AI reverse-engineering tool that uses Ghidra and LLMs to reconstruct and validate C/C++ functions from compiled binaries. ### Key features: - Reconstructs and validates C/C++ functions from compiled binaries. - Combines independent reverse and checker models. - Agentic evidence gathering and candidate build and test gates. - Structural verification and parity analysis. ### Tech stack: Python ### Why it matters: ReAgent is a powerful tool for reverse-engineering C/C++ code from compiled binaries, making it easier to understand and modify legacy systems.