Projects

Overview: This is a React.js library that converts text to speech using the Web Speech API, offering real-time text highlighting, dynamic controls, unlimited text input, and seamless speech playback. It supports multiple instances and auto-cleanup, making it easy to integrate into applications. Key features: - Converts text to speech using the Web Speech API - Highlights text in real time - Adjusts pitch, rate, volume, language, and voice programmatically - Automatically splits long text into chunks - Supports multiple speech instances - Automatically cancels speech on unmount - Provides error and event handling Tech stack: TypeScript Why it matters: This library simplifies integrating text-to-speech functionality into React applications, making it easy to create accessible and engaging user experiences without the need for external APIs or services.

Overview: This project provides an open-source offline speech recognition solution for Android using Mozilla's DeepSpeech, implemented in Termux. Key features: - Offline speech recognition - Uses Mozilla's DeepSpeech model - Compatible with Termux environment - Requires approximately 3GB of disk space during installation Tech stack: Shell Why it matters: This script offers a practical and offline alternative to traditional speech recognition software, making it useful for users without internet connectivity. Repository: T-vK/Termux-DeepSpeech GitHub stars: 86

### Overview: MFQ is a cutting-edge quantization and inference system designed to run large models efficiently on local hardware, optimizing memory, storage, and latency while maintaining model quality. ### Key Features: - **Fine-Grained Precision**: MFQ assigns precision to individual experts, projections, chunks, or neurons, ensuring higher precision where it's needed most. - **Mixed-Precision Formats**: It supports NINT scalar and NVQ vector formats, allowing for independent choices for each expert's Gate, Up, and Down projections. - **Native Runtimes**: The system executes mixed-precision formats directly, reducing reconstruction error and improving overall performance. - **Compact Weight Formats**: MFQ reduces weight residency, making it suitable for devices with limited memory, such as Metal and CUDA. ### Tech Stack: C++ ### Why It Matters: MFQ is crucial for developers who need to run large models on their existing hardware without compromising performance or model quality. By optimizing quantization and inference, it enables more efficient use of resources, leading to better model performance and faster deployment times.

Overview: This project is the official implementation of the paper "Align and Attend: Multimodal Summarization with Dual Contrastive Losses" presented at CVPR 2023. The repository contains the model, dataset, and training scripts for evaluating the model on various multimodal summarization tasks. Key features: - Model architecture for multimodal summarization - Training scripts for different datasets - Dataset structure and pre-processing instructions Tech stack: Python, PyTorch, Tensorboard, Rouge-Score, Scipy, ORTools, H5Py, YAML Why it matters: This project provides a comprehensive solution for multimodal summarization, demonstrating the effectiveness of the Align and Attend model in various applications.

Overview: This project is a comprehensive skill platform designed to streamline AI-assisted engineering workflows. It provides standardized, high-impact AI skills across multiple platforms, including Claude Code, GitHub Copilot, and 6 more. Key features: - Standardized AI skills across 18 universal platforms - Seamless integration with Claude Code, GitHub Copilot, and 6 other platforms - Automatic updates and new installs recommended by the installer - Supports both global and local installation methods Tech stack: JavaScript Why it matters: This platform is essential for developers looking to automate their AI workflows efficiently, ensuring consistent and high-impact results across various tools and platforms. Repository: ericgandrade/claude-superskills GitHub stars: 85 GitHub description: 18 Universal AI Skills for Claude Code, GitHub Copilot & 6 more platforms. Meta/orchestration, planning, research, and content — the focused core of the Superskills family.

Overview: This is a native Windows port of the NInfer engine, featuring Qwen3.8-27B with QUASAR and NInfer models, and MTP/DFlash2 with vision and 262,144 context. The engine is designed for maximum single-GPU inference performance. Key features: - Qwen3.8-27B with QUASAR and NInfer models - MTP/DFlash2 with vision and 262,144 context - Eight measured launchers over four Qwen3.8-27B artifact lines, all vision-capable at the full 262,144-token context - DFlash2 + MTP4, DFlash2 + MTP4, MTP4, DFlash2, DFlash2 + MTP4 - Upstream publishes artifacts for other checkpoints, which this port neither ships nor measures - Upstream requires a v3 container, which this port neither ships nor measures Tech stack: C++ Why it matters: This native Windows port of NInfer engine offers maximum single-GPU inference performance, making it ideal for developers looking to optimize their models for Windows environments.

Overview: This repository contains Jupyter Notebooks for experimenting with negative prompting in Stable Diffusion 2.0, including a personal notebook for inference, a grid generator, and a synthetic image generator for textual inversion. Key features: - SD 2.0 Base notebook - 3x3 Grid notebook - Wrong Image Generator notebook Tech stack: Jupyter Notebook Why it matters: These notebooks provide a practical way to experiment with negative prompting in Stable Diffusion 2.0, facilitating the development and testing of new textual inversion tokens.

Overview: This project is a local server implementation of the OpenAI Whisper API, running on FastAPI, designed for companies behind proxies or security firewalls. Key features: - Emulates OpenAI Whisper API for speech-to-text functionality - Compatible with OpenAI Whisper API and Audio Generation API - Supports various output formats, including text and audio - Compatible with both Python and Docker Tech stack: Python Why it matters: This server is essential for companies that need to use the Whisper API but are behind firewalls or proxies, providing a local alternative that can be accessed without restrictions. Repository: morioka/tiny-openai-whisper-api GitHub stars: 86

Overview: Minipilot is a GenAI-assisted chatbot that leverages Redis for vector database storage and vectorized RAG using Redis' vector search capabilities. It allows users to train the chatbot with their data, enabling natural language interactions. Key features: - Upload CSV data - Create an index - Edit system and user prompts - Ask questions in natural language Tech stack: Redis Stack as a vector database for storing and vectorizing data. Why it matters: Minipilot provides a practical solution for integrating AI into chatbots, making it accessible and efficient for various applications, especially those requiring vector database storage and natural language processing. Repository: redis/minipilot Primary language: HTML GitHub stars: 86 GitHub description: MiniPilot is a GenAI-assisted chatbot backed by Redis. Chat with your documents.

Overview: This project aims to decompose convolutional layers in a model to reduce the number of floating-point operations and parameters, specifically focusing on AlexNet, VGG, and ResNet50. Key features: - Decomposes convolutional layers in AlexNet, VGG, and ResNet50 models. - Uses CP and Tucker decomposition methods to optimize model performance. - Implements the decomposition in PyTorch without switching to numpy. Tech stack: PyTorch. Why it matters: The project showcases the effectiveness of CP and Tucker decomposition in reducing computational complexity and parameter count without compromising model accuracy, making it valuable for deep learning model optimization. Repository: ruihangdu/Decompose-CNN Primary language: Python GitHub stars: 86 GitHub description: CP and Tucker decomposition for Convolutional Neural Networks Usage: The script can decompose models and fine-tune parameters, with options for dataset path, decomposition type, model checkpoint, and state dictionary. Results: The project demonstrates that Tucker decomposition leads to lower accuracy loss compared to CP decomposition, showing significant improvements in model performance.

Overview: This project is a Unity plugin that allows developers to generate images, textures, and 3D assets directly within the Unity Editor using Scenario's AI models. Key features: - Direct integration with the Unity Editor - Seamless access to Scenario's AI-powered features - Simplifies asset creation and management - Compatible with C# and Unity Tech stack: C# Why it matters: This plugin streamlines asset creation, making it easier for Unity developers to generate high-quality assets directly within the Unity environment, enhancing productivity and creativity.

### Overview: SciTeX is a Python library designed to streamline the reproducible research process, enabling scientists to easily integrate and orchestrate various scientific tools and workflows. ### Key Features: - **Unified Toolkit**: Provides 73 modules under one namespace, accessible via Python API, CLI, and MCP (Multi-Component Plugin) for seamless integration. - **Cryptographic Verification**: Clew builds SHA-256 hash-chain DAGs to ensure every claim in a manuscript is linked to its source data. - **323 MCP Tools**: Enables AI agents to run statistics, create figures, search literature, and compile manuscripts through structured tool calls. - **Custom App Creation**: Researchers can create custom apps using the scitex-app SDK and share them on SciTeX Cloud. - **Open and Self-Hostable**: All SciTeX packages are AGPL-3.0 licensed, and the full ecosystem runs on your own hardware or SciTeX Cloud, which is self-hostable. ### Tech Stack: - Python ### Why It Matters: SciTeX simplifies the complex process of scientific research by providing a unified, cryptographic-verified, and self-hostable platform for data integration and reproducibility. It enables researchers to streamline their workflows, ensuring every step is traceable and verifiable, and helps them avoid vendor lock-in, making it an essential tool for modern research. ### Repository: - **Repository**: scitex-ai/scitex-python - **Stars**: 86 - **Documentation**: scitex-python.readthedocs.io - **License**: AG

### Overview: Multimodal entity linking (MEL) aims to utilize multimodal information to map mentions to corresponding entities defined in knowledge bases. We release three MEL datasets: Weibo-MEL, Wikidata-MEL, and Richpedia-MEL, each containing 25,602, 18,880, and 17,806 samples from social media, encyclopedia, and multimodal knowledge graphs, respectively. The datasets are constructed through five stages: multimodal information extraction, mention extraction, entity extraction, triple construction, and dataset construction. The provided datasets and construction approach demonstrate the usability and distinguishability of baseline models. ### Key features: - **Datasets**: Weibo-MEL, Wikidata-MEL, and Richpedia-MEL, each containing specific sample counts from social media, encyclopedia, and multimodal knowledge graphs. - **Construction Approach**: A five-stage approach including multimodal information extraction, mention extraction, entity extraction, triple construction, and dataset construction. - **Visualization**: Visual information is available for download, including images and textual descriptions. ### Tech stack: Python, JavaScript, and various machine learning libraries. ### Why it matters: The datasets and construction approach provide a valuable benchmark for evaluating the performance of MEL models, ensuring that researchers can accurately measure the effectiveness of their models and improve their algorithms.

Overview: MJ Studio 是一个集成了多种图像生成模型的多模型 AI 绘图工作台，支持包括 Midjourney、Gemini、Flux、DALL-E、GPT-4o、Grok、通义万相等在内的主流图像生成模型。它提供了一个轻量级的部署方式，无需依赖 Redis 或 PostgreSQL，仅需 3 条命令即可运行，支持 SQLite 数据库和本地图片存储。 Key features: - 绘图模块：支持多种主流图像生成模型。 - 视频模块：支持 Sora、即梦、Veo、Grok Video 等视频生成模型，支持异步轮询生成和视频预览下载。 - 对话模块：支持多助手管理、流式输出、对话分叉、智能压缩和消息重放。 Tech stack: TypeScript Why it matters: MJ Studio 是一个轻量级的多模型 AI 绘图工作台，支持多种主流图像生成模型，提供了一个轻量级的部署方式，无需依赖 Redis 或 PostgreSQL，仅需 3 条命令即可运行，支持 SQLite 数据库和本地图片存储，特别适合需要高效部署和隐私保护的场景。

### Overview: The STACKIT RAG Template is a comprehensive project that provides a foundational template for building AI chatbots and document management systems using Retrieval-Augmented Generation (RAG). It includes essential libraries and APIs for vector search, FastAPI, and more. ### Key Features: - **Vector Search**: Utilizes vector search to enhance the accuracy of document retrieval. - **FastAPI**: A lightweight and efficient framework for building APIs. - **RAG-API Libraries**: Includes libraries for RAG, RAG-Extractor, and RAG-Admin, providing a complete toolkit for AI applications. ### Tech Stack: - Python ### Why It Matters: This project is crucial for developers looking to build intelligent AI chatbots and document management systems. It simplifies the integration of vector search and FastAPI, making it easier to implement RAG technology in various applications. The libraries included are designed to be robust and efficient, ensuring that developers can focus on the core functionality of their projects without worrying about the underlying infrastructure.

### Overview: Voltip is a dictation app designed for seamless dictation and text input across Windows, Linux, macOS, and Android. It leverages cloud and on-device speech recognition, AI polish, and a comprehensive feature set including live preview, voice editing, and long recording capabilities. The app supports multiple languages and provides a user-friendly interface. ### Key features: - **Push-to-talk Hotkey**: Use `Ctrl+Alt+Space` by default or a single key like `Right Ctrl` or a mouse side button. - **Recognition from Cloud or Device**: Recognize speech from cloud providers or on-device with Qwen3-ASR, SenseVoice, and Sherpa-Onnx. - **Live Preview**: Real-time transcription while you speak. - **Clean-up by AI**: Polish text with LLMs (Google AI Studio, DeepSeek, Kimi, Zhipu GLM, Volcengine Ark, Tencent Hunyuan, Xiaomi MiMo, OpenRouter, and Ollama). - **Voice Edit**: Select text, hold `Ctrl+Alt+E`, and use voice commands to refine the text. - **Scenes**: The app's focus style, output mode, language, and extra instructions for each take. - **Long Recordings**: Record up to two hours of audio, transcribe in segments, and export as SRT subtitles or plain text. ### Tech stack: Rust The app is built using Rust, a language known for its safety and performance. ### Why it matters: Voltip simplifies dictation and text input, making it

Overview: AmazingDraw 是一个本地 ComfyUI 抽卡引擎，旨在解决不会写提示词的问题。内置场景库和角色库，只需一句话即可抽卡，无需自己攒 prompt。 Key features: - 自动解压到固定目录 - 支持多种 AI Agent，如 Cursor、Claude Code、OpenClaw、Cherry Studio - 一句话抽卡，支持随机抽卡和指定场景库和角色库 - 支持 macOS 和 Windows 平台 - WebUI 可选，需安装特定的 Python 版本 Tech stack: Python Why it matters: 解决不会写提示词的问题，提供了一种快速、简便的抽卡方式，无需手动编写复杂的 prompt。 Repository: AmazingDraw/AmazingDraw GitHub stars: 84 GitHub description: 本地 ComfyUI 抽卡引擎，解决不会写提示词的问题。

### Overview: ReQFlow is an efficient and high-quality protein backbone generation model that achieves state-of-the-art performance while significantly reducing sampling steps and inference time. It is particularly effective for generating protein backbones of moderate length. ### Key Features: - **Efficient Sampling**: ReQFlow requires fewer sampling steps compared to existing methods, making it faster and more efficient. - **High-Quality Backbones**: The model generates high-quality protein backbones, demonstrating its effectiveness. - **Significantly Faster Inference**: Compared to RFDiffusion and Genie2, ReQFlow is 37× faster and 62× faster in generating a backbone of length 300. ### Tech Stack: - Python ### Why It Matters: ReQFlow is a groundbreaking model that addresses the current limitations of protein backbone generation by providing both high performance and efficiency. Its ability to generate high-quality backbones quickly makes it a valuable tool for researchers and practitioners in the field of protein structure prediction and design.

Overview: `tracedistill` is a library that distills teacher chains-of-thought into a LoRA adapter, enabling models to rederive answers without running any code. This approach is applicable to tasks where the grader cannot run the code, making it ideal for the NVIDIA Nemotron Model Reasoning Challenge. Key features: - Distills teacher chains-of-thought into a LoRA adapter - Preserves the medal-winning code from the competition - Supports a strict boxed-answer format - Uses two-phase training to handle hard and easy problems - Implements type-stratified batching to ensure balanced gradients - Uses architecture-aware LoRA to cover the entire model architecture Tech stack: Python Why it matters: This library provides a practical solution for tasks where the grader cannot run the code, making it ideal for the NVIDIA Nemotron Model Reasoning Challenge.

### Overview: This repository maintains a collection of important papers on knowledge distillation, focusing on various aspects such as pioneering papers, survey papers, and advancements in knowledge distillation techniques applied to diffusion models. ### Key Features: - **Pioneering Papers**: Includes seminal works on model compression and deep learning optimization. - **Survey Papers**: Offers comprehensive reviews and outlooks on knowledge distillation and related fields. - **Distillation Meets Diffusion Models**: Highlights recent advancements in accelerating sampling of diffusion models using knowledge distillation techniques. ### Tech Stack: Python and TensorFlow/PyTorch are primarily used for implementing and experimenting with knowledge distillation algorithms. ### Why It Matters: Knowledge distillation is a powerful technique that enables efficient model compression while maintaining high accuracy. By applying this technique to diffusion models, researchers have achieved significant improvements in sampling speed and efficiency, making it an essential tool in the field of computer vision and machine learning.