Projects

Overview: NeuroPrompt Studio is a comprehensive repository designed for advanced language model orchestration, focusing on pattern engineering and performance optimization. Key features: - **Pattern Library**: Atomic prompt structures for zero-shot, few-shot, chain-of-thought, tree-of-thought, structured output, and reflection patterns. - **Workflow Blueprints**: Multi-step interaction sequences for complex tasks, including document analysis, code generation, and research synthesis. - **Evaluation Suite**: Metrics and testing harnesses for prompt performance, including consistency scores, hallucination detection, and latency benchmarks. Tech stack: HTML Why it matters: NeuroPrompt Studio transforms prompt engineering from an art into a systematic science, providing practical frameworks and benchmarks for developers, researchers, and creative technologists to enhance the performance and reliability of language models. Repository: thenicolas1894/awesome-claude-fable-5-prompt-vault GitHub stars: 140

### Overview: CopywriterProAI is the world's first open-source AI content writing platform designed to help users create SEO-friendly blog posts, ad copy for social media, website landing pages, and more. Empowering users with the power of AI, it streamlines the content creation process, ensuring high-quality output without the need for manual labor. ### Key Features: - **SEO-Friendly Content Creation**: Automatically optimize blog posts and ad copy for better search engine rankings. - **Customizable Templates**: Choose from various templates to match the style and tone of your content. - **Real-time Collaboration**: Work on content together with team members in real-time. - **Customizable AI Models**: Tailor the AI to your specific needs and preferences. - **Integration with Popular Tools**: Seamless integration with popular tools like Google Analytics, Hotjar, and more. ### Tech Stack: Node v12.0+, Yarn, MongoDB Database ### Why it Matters: CopywriterProAI is a game-changer in the field of AI content writing, offering a comprehensive solution for businesses looking to improve their online presence. Its user-friendly interface and seamless integration with popular tools make it an indispensable tool for marketers, content creators, and anyone seeking to boost their online visibility.

### Overview: Ground-A-Video is a zero-shot Grounded Video Editing framework that integrates spatially-continuous and spatially-discrete conditions for multi-attribute video editing. It achieves temporally consistent multi-attribute editing in a training-free manner, preserving non-target regions, and surpassing baseline methods in edit-accuracy and frame consistency. ### Key Features: - **Cross-Frame Gated Attention**: Incorporates groundings information into latent representations temporally consistently. - **Modulated Cross-Attention**: Enhances editing accuracy by modulating cross-attention mechanisms. - **Optical Flow Guided Inverted Latents Smoothing**: Preserves input video regions that should remain intact. ### Tech Stack: PyTorch ### Why it Matters: Ground-A-Video is a significant advancement in the field of video editing, offering a training-free approach to multi-attribute video editing, thereby overcoming limitations of existing methods.

Overview: This repository contains the MXNet implementation of the RNN Transducer (Graves 2012), a sequence transduction model for end-to-end speech recognition. The repository includes various scripts and models for training, evaluation, and feature extraction, making it a comprehensive resource for researchers and developers in the field of speech recognition. Key features: - RNN Transducer model for end-to-end speech recognition - Joint model for RNNT (Sequence Transduction with Recurrent Neural Networks) - Support for both symbol and gluon data processing - Implementation of RNNT loss function on MXNet - Training scripts for RNNT, CTC, and attention models - Evaluation scripts for greedy and beam search decoding Tech stack: MXNet Why it matters: This repository provides a practical implementation of the RNN Transducer model, which is essential for researchers and developers looking to advance the field of speech recognition. The inclusion of joint models and support for various data processing methods makes it a valuable resource for both academic research and practical applications. Repository: HawkAaron/RNN-Transducer Primary language: Python GitHub stars: 140 GitHub description: MXNet implementation of RNN Transducer (Graves 2012): Sequence Transduction with Recurrent Neural Networks

### Overview: This repository is the application detailed in the "Learning Blazor: Build Single-Page Apps with WebAssembly and C#" O'Reilly Media book by David Pine. It is a Blazor WebAssembly app deployed to Azure as a Static Web app, targeting .NET 6 and packed with C# 10 features. ### Key Features: - Blazor Third-Party A - Blazor Third-Party B - Blazor Third-Party C ### Tech Stack: C# ### Why it Matters: This repository provides a practical guide for developers to learn how to use Blazor WebAssembly to develop next-generation web experiences. It is built on top of ASP.NET Core and targets the future of .NET single-page applications (SPA) investments. The app is designed to enhance C# skills and demonstrate the capabilities of C# 10 features.

### Overview: The Brain Latent Progression (BrLP) method is an innovative approach to enhancing spatiotemporal disease progression models by leveraging latent diffusion and prior knowledge. This project presents a comprehensive implementation of BrLP, which has been recognized with a top-tier award at the MICCAI conference. ### Key features: - **BrLP Method**: A novel technique for improving disease progression models by incorporating latent diffusion and prior knowledge. - **Latent Diffusion**: Utilizes latent diffusion to capture complex temporal patterns in disease progression. - **Prior Knowledge**: Incorporates prior information to refine predictions and improve model accuracy. - **Spatiotemporal Models**: The method is specifically designed for spatiotemporal disease progression, making it highly relevant for medical imaging applications. ### Tech stack: - **Python**: The primary programming language used for implementing the BrLP method. - **TensorFlow/Keras**: Used for training and evaluating the neural network components of the model. ### Why it matters: - **Enhanced Accuracy**: BrLP significantly improves the accuracy of spatiotemporal disease progression models by incorporating prior knowledge and latent diffusion. - **Reproducibility**: The project provides detailed documentation on data preparation and reproducibility, ensuring that the results can be easily replicated. - **Medical Applications**: The method is particularly useful in medical imaging, where precise disease progression tracking is crucial for patient care and treatment planning. ### Repository: - **Repository**: LemuelPuglisi/BrLP - **GitHub Stars**: 140 - **Description**: Official

Overview: OpenSTT is an open source project aiming to develop an open source speech-to-text model. Key features: - Evaluation of speech-to-text model approaches and frameworks - Development of an open source speech-to-text model Tech stack: Python, TensorFlow, Kaldi Why it matters: This project aims to contribute to the open source community by providing an open source speech-to-text model, which can be beneficial for research, development, and collaboration in the field of speech recognition.

### Overview: HalluQA is a comprehensive dataset designed to evaluate the hallucination capabilities of Chinese large language models. It includes 450 adversarial questions covering various domains, ensuring a realistic assessment of model performance. ### Key features: - Adversarial questions crafted to induce hallucinations in Chinese large language models. - Detailed experimental results for multiple Chinese large language models. - Multiple-choice task for evaluating model performance in a specific context. ### Tech stack: Python ### Why it matters: HalluQA provides a robust framework for evaluating the hallucination capabilities of Chinese large language models, ensuring that models accurately represent the knowledge they have been trained on.

Overview: Sin3DM is a project that focuses on learning a diffusion model from a single 3D textured shape, aiming to generate new textured meshes through diffusion techniques. Key features: - Single 3D textured shape learning - Diffusion model training and generation - Pre-trained 3D shape classifier for evaluation Tech stack: Python Why it matters: Sin3DM demonstrates the potential of learning diffusion models from a single textured shape, showcasing their ability to generate realistic 3D textures, which could have applications in areas like 3D modeling, art generation, and more. Repository: Sin3DM/Sin3DM GitHub stars: 140 GitHub description: single 3D shape diffusion model

Overview: This repository contains the official implementation of our paper "EditWorld: Simulating World Dynamics for Instruction-Following Image Editing" presented at ACM Multimedia 2025 Datasets Track. It introduces a new task called "world-instructed image editing" and provides a new image editing dataset with world instructions. Key features: - World-instructed image editing task - New image editing dataset with world instructions - Proposed new post-edit method for world-instructed image editing Tech stack: Python Why it matters: This work introduces a novel approach to image editing by incorporating world dynamics, making it a significant contribution to the field of image editing and instruction-following techniques.

### Key features: - Concept Representation and Extraction - Concept Activation Detection - Activation Vector Steering ### Tech stack: Node.js ### Why it matters: This repository serves as a curated list of resources and papers related to activation engineering, which is an emerging field focused on understanding and manipulating model activations to enhance model interpretability and control.

### Overview: This repository is a curated list of open-source repositories for AI engineers, focusing on various aspects of AI infrastructure, workload management, copilot development, dataset engineering, and more. ### Key Features: - **AI Gateway**: A collection of repositories for building and managing AI gateways. - **AI Workload Manager**: Repositories for managing and optimizing workload in AI environments. - **Copilot Development**: Tools and frameworks for developing AI copilots. - **Dataset Engineering**: Repositories for creating, managing, and engineering datasets. - **Evaluation**: Tools and frameworks for evaluating AI models. - **Fine Tuning**: Repositories for fine-tuning and optimizing AI models. - **Function Calling**: Libraries and tools for calling AI functions. - **Graph RAG**: Repositories for graph-based reasoning and artificial intelligence. - **Guardrails**: Tools for designing and enforcing AI policies. - **Local Model Inference**: Repositories for running AI models locally. - **LLM Agent Framework**: Frameworks and tools for building AI agents. - **Model Serving**: Repositories for deploying and serving AI models. - **Observability**: Tools and frameworks for monitoring and observing AI systems. - **Pre Training**: Repositories for pre-training AI models. - **Prompt Engineering**: Tools and frameworks for crafting effective AI prompts. - **RAG Framework**: Repositories for reasoning and argument generation. - **Structured Extraction**: Tools for extracting structured information from text. - **Structured Generation**: Repositories for generating structured outputs. - **Vector DB**: Repositories for working with vector databases in AI

Overview: Spectrum is an adaptive spectral feature forecasting method designed for diffusion sampling acceleration, enabling global, long-range feature reuse with controlled error. Key features: - Adaptive spectral feature forecasting - Tightly controlled error - Approximates latent features with Chebyshev polynomials - Theoretical error bound that does not compound with step size Tech stack: Python Why it matters: Spectrum significantly accelerates diffusion sampling by up to 4.79x on FLUX and 4.67x on Wan2.1-14B, while maintaining higher sample quality compared to existing methods.

### Overview: This project introduces a feed-forward VQGAN-CLIP model designed to eliminate the need for optimizing the latent space of VQGAN for each input prompt. By training a model that directly maps text prompts to VQGAN latent spaces, it aims to simplify the process and potentially improve efficiency and performance. ### Key Features: - **Text-to-Image Generation**: The model takes text prompts as input and outputs the corresponding VQGAN latent space. - **Latent Space Mapping**: It transforms the latent space into RGB images, effectively bypassing the need for optimizing the latent space for each prompt. - **Loss Function**: The model minimizes the distance between CLIP-generated image features and the input text features, ensuring consistency with CLIP's image generation capabilities. - **Diversity Loss**: An additional diversity loss is introduced to enhance the variety of generated images given the same prompt. ### Tech Stack: - **Python**: The primary programming language used for developing and training the model. ### Why It Matters: This model simplifies the process of generating images from text prompts, potentially reducing the need for complex optimization and improving performance. It leverages the strengths of CLIP and VQGAN to create high-quality images from textual descriptions, making it particularly useful in applications requiring high-resolution image generation from text inputs. ### Repository: - **Repository**: mehdidc/feed_forward_vqgan_clip - **GitHub Stars**: 140 - **Primary Language**: Python - **GitHub Description**: Feed forward VQGAN-CLIP model, eliminating the need for optimizing

### Overview: Off Grid AI is a local-first AI runtime for macOS and Windows, providing an on-device personal AI assistant. It includes an OpenAI-compatible API, an always-on private layer, and chat, image generation, voice functionality, and file integration. The app is designed to run entirely on your machine, ensuring no cloud or account usage. ### Key features: - **Chat**: Supports text, vision, and reasoning. - **Image Generation**: Generates on-device images. - **Voice**: Offers live voice input and output. - **File Integration**: Connects tools with local models. - **Pro Layer**: Provides local, memory, reflection, and action capabilities. ### Tech stack: TypeScript ### Why it matters: Off Grid AI is designed to be a private, on-device AI assistant, eliminating the need for cloud storage, accounts, or API keys. It runs entirely on your machine, ensuring your conversations, files, and models stay local and secure. The Pro layer adds advanced features like seeing, remembering, reflecting, and acting on-device, making it a powerful tool for personal AI. ### GitHub stars: 139 ### Repository: off-grid-ai/OGAD ### Primary language: TypeScript ### GitHub description: Your private, on-device personal AI assistant for macOS and Windows. Chat, create, search files, and connect tools with local models. Includes an OpenAI-compatible API. Opt-in Pro sees, remembers, reflects, and acts with your approval. No cloud or account.

### Overview: Art 是一个开源的、一站式 AI 应用开发平台，专为 Java 技术栈团队和企业设计。它旨在通过 Spring Boot 和 Spring Cloud 等成熟框架，帮助 Java 开发者在熟悉的环境中构建、部署和管理生产级的 AI 应用。 ### Key features: - **可视化 AI 工作流**: 拖拽式画布编排语言模型、工具集、代码块和知识库，构建复杂 AI 应用。 - **新一代 RAG 引擎**: 融合知识图谱与向量数据库，提供深层次的逻辑推理和关系挖掘。 - **无缝拥抱 Spring Boot 和 Spring Cloud**: 将企业服务和数据源封装成自定义节点，释放现有 IT 资产潜力。 - **微服务架构**: 支持统一接入和管理多种模型服务商，基于高内聚、低耦合的微服务架构设计。 - **完全开源**: 遵循 Apache 2.0 协议，所有代码开源，社区驱动，透明可信。 ### Tech stack: Java ### Why it matters: Art 是为 Java 技术栈团队和企业量身定制的 AI 应用开发平台，通过 Spring Boot 和 Spring Cloud 等成熟框架，提供一个稳定、高效且易于扩展的环境，让开发者在熟悉的 Java 环境中快速构建、部署和管理 AI 应用。

Overview: Libratory is a free book and audiobook laboratory that converts your existing PDFs into chapter-marked audiobooks, read-along books, and bilingual editions. It supports Apple Silicon Macs, Linux, and Docker containers, offering offline-first functionality and local AI cleanup and translations. Key features: - Converts PDFs into audiobooks, read-along books, and bilingual editions. - Runs on your own machine, including Apple Silicon Macs, Linux, and Docker containers. - Supports local TTS, AI cleanup, and translations. - Allows editing, re-narrating, and excluding chapters. - Supports translation and rewriting of chapters. - Provides an assistant for searching and running the library. - Offers digest books, read-along narration, and two-language support. - Allows exporting selected chapters as PDF, EPUB, or bilingual EPUB. - Supports library organization with nested folders, drag & drop, and cross-folder search. Tech stack: TypeScript Why it matters: Libratory simplifies the process of converting and organizing your existing PDFs into audiobooks and bilingual editions, making it easier to access and share knowledge offline.

### Overview: Seismic is a fast and lightweight search engine for learned sparse embeddings, written in Rust with Python bindings. It indexes sparse vector collections and retrieves results in microseconds with near-exact accuracy. ### Key Features: - **Multiple index variants**: Standard (`SeismicIndex`), compressed (`SeismicIndexDotPacking8`), and large vocabulary (`SeismicIndexLV`). - **RAG-ready**: Build the index with `load_content=True` and retrieve document texts alongside scores. - **Python & Rust APIs**: Use from Python via `pyseismic-lsr` or integrate directly in Rust via `cargo add seismic`. - **Parallel batch search**: Multi-threaded query processing via `batch_search`. ### Tech Stack: Rust ### Why it Matters: Seismic is designed for efficient and accurate retrieval of sparse embeddings, making it ideal for applications requiring fast and precise search capabilities. Its lightweight implementation and Python/Rust bindings make it accessible and versatile for a wide range of use cases.

Overview: Pansori is a program designed to create an automatic speech recognition (ASR) corpus from online videos containing audio and subtitle data. It consists of four pipeline stages: ingest, align, transform, and validate. Key features: - Downloads audio and subtitle streams from online videos. - Aligns subtitles with audio based on timing information. - Processes audio and subtitle data through specific transformations. - Validates the corpus by filtering out inaccurate audio and subtitle pairs. Tech stack: Python Why it matters: Pansori simplifies the process of creating an ASR corpus from online video content, making it accessible and efficient for developers and researchers. The use of Google Cloud Speech-to-Text API for validation ensures high-quality results across multiple languages, addressing the need for custom ASR models in many languages.

Overview: This project demonstrates how to fine-tune IndexTTS to generate speech with additional special tags, such as "<GIGGLES>", enabling the synthesis of non-textual elements like laughter. Key features: - Fine-tuning of IndexTTS's text Tokenizer (BPE) and AR part (GPT2) - Support for addtional special tags like "<GIGGLES>" - Fine-tuning of a dataset using [🤗 MrDragonFox/Elise](https://huggingface.co/datasets/MrDragonFox/Elise) - Fine-tuning experiment results example Tech stack: Jupyter Notebook Why it matters: This project serves as a demonstration of fine-tuning IndexTTS, showcasing how to add special tags to the text tokenizer and generate speech with them, making it useful for generating non-textual elements like laughter. Repository: yrom/finetune-index-tts Primary language: Jupyter Notebook GitHub stars: 140 GitHub description: IndexTTS Fine-tuning notebooks License: MIT License