Projects

### Overview: This repository contains an implementation of the Reparameterized Discrete Diffusion Model, a novel approach for text generation using discrete diffusion processes. The model is designed to be flexible and reusable, with a self-contained library that supports various types of discrete diffusion models. ### Key Features: - **Discrete Diffusion Models**: Implements various types of discrete diffusion models, including vanilla and reparameterized versions. - **Self-contained Library**: Provides a library `discrete_diffusion` for general use, supporting different diffusion processes. - **FairSeq Dependencies**: Uses the `fairseq` library for installation, ensuring compatibility with specific versions of Python, PyTorch, and CUDA. ### Tech Stack: Python ### Why it Matters: This model offers a flexible and efficient approach to text generation, leveraging the strengths of discrete diffusion and fairseq libraries to create a robust and versatile solution.

### Overview: Kumoss is a powerful tool that converts natural-language requests into validated Terraform/OpenTofu pull requests, ensuring compliance and maintaining the integrity of infrastructure operations. ### Key Features: - **Natural-Language Requests**: Converts human-readable requests into structured code. - **Compliance and Validation**: Ensures compliance with policies and standards. - **Drift Correction**: Corrects drift issues automatically. - **Policy Gates**: Enforces rules and gates to prevent high-risk changes. - **Human Review**: Keeps humans in control of high-impact changes. - **Multicloud Support**: Works across multiple cloud environments. ### Tech Stack: Python ### Why It Matters: Kumoss simplifies infrastructure operations by providing a platform for developers to create, update, and maintain their own infrastructure, ensuring compliance and reducing the need for platform team involvement. ### Repository: InditexTech/kumoss ### GitHub Stars: 95 ### GitHub Description: Kumoss turns natural-language requests into compliant, plan-validated Terraform/OpenTofu pull requests - with drift correction, policy gates, and human review. Multicloud, bring-your-own LLM.

### Overview: MicroCore is a Python library designed to facilitate interaction with large language models (LLMs) and other AI services, making it easy to switch between different models and services without altering your application code. It provides a minimalistic foundation for AI applications, separating business logic from implementation details. ### Key Features: - **LLM Communication**: Smooth interaction with various LLMs using the Minimal Comprehensible Protocol (MCP). - **Vector Databases**: Integration with vector databases for semantic search. - **Configuration**: Minimal configuration options for seamless integration with different models and services. ### Tech Stack: - Python 3.10 / 3.11 / 3.12 / 3.13 / 3.14 ### Why It Matters: MicroCore simplifies the integration of AI services into your applications, allowing you to focus on the core functionality of your project while keeping the implementation details separate. This approach makes your application more maintainable and easier to update as new models or services become available. ### Repository: The repository is hosted on GitHub at [Nayjest/ai-microcore](https://github.com/Nayjest/ai-microcore). It has received 108 GitHub stars and is maintained by the same developer. ### Installation: - **PyPI Package**: Install using `pip install ai-microcore`. - **Local Installation**: Copy the `microcore` folder to your project's root directory and remove the original folder. ### Links: - [API Reference](https://ai-microcore.github.io/api-reference

Overview: AICLIPS Studio is an open-source AI video clipping generator designed for the production-ready Next.js SaaS ecosystem. It extracts viral shorts, Reels, and TikToks from long-form YouTube videos, featuring features like seamless YouTube video extraction, asynchronous AI highlight detection, and secure credit charging. Key features: - Seamless YouTube video extraction - AI highlight engine for engaging moments - Custom aspect ratio tuning - Secure credit system with real-time cost calculation - Historical archive for logged-in users - Asynchronous webhooks for database updates Tech stack: JavaScript, Next.js, Stripe, Prisma, PostgreSQL, OAuth, and TypeScript. Why it matters: AICLIPS Studio offers a production-ready, fully-integrated solution for extracting viral content from YouTube videos, making it ideal for developers looking to automate the process of generating popular short clips. Repository: SamurAIGPT/ai-clipping-generator GitHub stars: 107

### Overview: This repository is a curated list of awesome resources related to OpenAI's Whisper model, including general resources, API-ready implementations, playgrounds, and various model variants and applications. ### Key Features: - General Resources: Introducing Whisper, Whisper Paper, Whisper Code, Introducing ChatGPT and Whisper APIs. - API Ready / Playground / Demo: whisperx, WHISPER+, Fine-Tuned Whisper API, openai/whisper – Run with an API on Replicate, Whisper Playground, Web Whisper, Uttera, Whipscribe, Whisper Webui. - Model Variants: whisper-timestamped, whisper.cpp, Faster Whisper, whisper-jax, whisper.tflite, OpenAI Whisper, whisper_onnx, whisper-export, whisper-openvino. - Applications: React hook for OpenAI Whisper, Subtitles generation tool, Whisper as a Service, WhisperX, stable-ts, buzz, whispering, whisper-youtube, Speaker Identification, AutoCut, Subtitle Edit. ### Tech Stack: Python, JavaScript, C++, C++, Hugging Face, TensorFlow, ONNX, OpenVINO. ### Why it Matters: This list provides a comprehensive overview of tools and resources related to OpenAI's Whisper model, making it a valuable resource for developers and researchers looking to integrate Whisper into their projects.

Overview: MCP Tool Kit is a comprehensive toolkit designed for building high precision vertical AI agents using the Model Context Protocol (MCP) in Python. It includes automatic tool discovery and multiple transport options, making it easy to integrate with various environments. Key features: - Dynamic tool registration - Multiple transport options (stdio, SSE) - 120+ available tools - Easy configuration through a `config.yaml` file Tech stack: Python Why it matters: MCP Tool Kit simplifies the development process by providing a robust, automated solution for integrating vertical AI agents with the Model Context Protocol, making it suitable for both developers and businesses looking to enhance their AI capabilities. Repository: getfounded/mcp-tool-kit GitHub stars: 108

### Overview: Quantum Speech Processing with CNNs is a project that explores the application of quantum computing in speech recognition. This project aims to leverage quantum neural networks to enhance the accuracy and efficiency of automatic speech recognition (ASR) systems. ### Key Features: - **Quantum Convolutional Neural Network (QCNN):** Implements a quantum convolutional neural network for speech processing tasks. - **Speech Commands Dataset V1:** Utilizes the Google Speech Commands Dataset for training and testing. - **Pre-processed Features:** Provides pre-processed mel features and (2,2) convolution features for training. - **Training and Testing Data:** Includes 2000 pre-processed features for training and 500 for testing, achieving a test accuracy of 90.6%. ### Tech Stack: - **Python:** Main programming language. - **TensorFlow:** Framework for implementing the QCNN. ### Why It Matters: - **Enhanced Accuracy:** Quantum computing can potentially provide a significant boost in speech recognition accuracy. - **Efficiency:** Quantum algorithms can process large datasets more efficiently than classical algorithms. ### Repository: - **Repository:** huckiyang/QuantumSpeech-QCNN - **Stars:** 108 - **Description:** IEEE ICASSP 21 - Quantum Convolution Neural Networks for Speech Processing and Automatic Speech Recognition. ### Environment: - **TensorFlow:** Installed via conda and pip. ### Dataset: - **Speech Commands Dataset V1:** Used for training and testing. - **Pre-processed Features:** Provided for training,

Overview: This repository contains the code for the book "The Developer's Guide to AI - A Field Guide for the Working Developer." Key features: - Pretrained AI Models - Project Structure - Prerequisites - Getting Started - Parts Overview: Part 1: Getting Started with AI, Part 2: Prompt Engineering, Part 3: Vector Databases and RAG, Part 4: Adapting Models to Real-World Tasks, Part 5: Building Agentic Systems - Creating the README Tech stack: Python Why it matters: This book provides practical guidance for developers looking to integrate and customize large language models and other pretrained AI models into their applications, avoiding the pitfalls of AI prototypes and focusing on real-world solutions. Repository: jorshali/developers-guide-to-ai GitHub stars: 108 GitHub description: The Developer's Guide to AI - A Field Guide for the Working Developer

### Overview: AutoPenBench is a penetration test benchmark designed for generative agents, presented in the paper "AutoPenBench: Benchmarking Generative Agents for Penetration Testing". It includes instructions for installing, developing, and testing new vulnerable containers for the benchmark. ### Key Features: - **Benchmarking Generative Agents for Penetration Testing**: This repository contains the codes for evaluating generative agents in penetration testing scenarios. - **Instructions for Installation, Development, and Testing**: Detailed instructions are provided to install, develop, and test new vulnerable containers for the benchmark. - **Supported Tasks**: The benchmark includes tasks related to access control, web security, network security, and cryptography for both in-vitro and real-world scenarios. ### Tech Stack: Python ### Why It Matters: AutoPenBench provides a standardized benchmark for evaluating generative agents in penetration testing scenarios, ensuring fair and comparable performance across different models and agents. This benchmark is crucial for researchers and developers to evaluate the capabilities and limitations of generative agents in real-world cybersecurity tasks. ### Repository: - **GitHub Stars**: 106 - **Repository**: lucagioacchini/auto-pen-bench - **Primary Language**: Python ### Installation: 1. Ensure `cmake` is installed. 2. Create and activate a virtual environment. 3. Install the requirements and set up the machines using `make install`. 4. To test one instance of the benchmark, refer to the example folder. ### How to Test and Evaluate an Agent: - Refer to the example notebook `instructor_agent

### Overview: SurrogateShield is a privacy-first tool designed to protect LLMs and RAG pipelines by automatically redacting and reconstructing sensitive PII data in real-time. It ensures that personal information is masked before it reaches the LLM, thereby enhancing privacy and security. ### Key Features: - **Layered PII Detection**: SurrogateShield uses regex patterns, DeBERTa PII tagger, and spaCy for location data to detect and redact PII. - **Realistic Surrogate Generation**: It generates PII data that mimics real-world patterns, including fake names, SSNs, and Bitcoin addresses. - **AES-256-GCM Encrypted ShadowMap**: Surrogate values are encrypted and stored in a ShadowMap, ensuring data integrity and confidentiality. ### Tech Stack: - Python ### Why It Matters: SurrogateShield provides a robust solution for privacy-preserving proxying, enhancing the security and privacy of large language models and retrieval-based systems.

### Overview: A scalable web crawler built with LLMs and Ray, utilizing Huggingface's BERT model to extract text embeddings. The embeddings are saved into a vector database for efficient clustering and retrieval, making it suitable for a rich-representation language model (RAG) system. ### Key Features: 1. **Scalable Web Crawler**: Crawl recursively the web, storing links and their corresponding text and embeddings. 2. **Large Language Model (LLM) Utilization**: Uses Huggingface's BERT model to extract text embeddings. 3. **Ray for Scalability**: Utilizes Ray to distribute the workload across multiple workers. 4. **Vector Database for Fast Retrieval**: Saves and retrieves entries using a vector database for efficient querying. ### Tech Stack: - Python - Ray - Huggingface ### Why It Matters: By saving the representations into a vector database, you can retrieve similar pages based on how close two vectors are, making it ideal for a rich-representation language model (RAG) system. ### Repository: - **Repository**: Aavache/LLMWebCrawler - **GitHub Stars**: 107 - **GitHub Description**: A Web Crawler based on LLMs implemented with Ray and Huggingface. The embeddings are saved into a vector database for fast clustering and retrieval. Use it for your RAG. ### Primary Language: - **Primary Language**: Python ### GitHub Description: - A Web Crawler based on LLMs implemented with Ray and Huggingface. The embeddings are saved into

Overview: This Python service orchestrates conversational agents in the Agent Landing Zone, leveraging Foundry Agent Service and Agent Framework for seamless execution and integration. Key features: - Orchestrates specialized agents for conversational experiences - Utilizes grounded knowledge, NL2SQL, and MCP tools for effective interaction - Coordinates with Cosmos DB for conversation storage Tech stack: Python Why it matters: This service streamlines the execution of conversational agents, enhancing the efficiency and effectiveness of conversational AI solutions.

Overview: Graph-CAD is a text-to-CAD framework that generates executable Blender code from natural language instructions, addressing the fragility of long-horizon CAD code generation by explicitly modeling hierarchical decomposition, geometric constraints, staged generation, and progressive curriculum learning. Key features: - Hierarchical and geometry-aware graph representation - Progressive curriculum learning for improved graph modeling - Modular three-stage inference pipeline - Pre-release codebase organized for public release - Comprehensive evaluation framework Tech stack: Python Why it matters: Graph-CAD reduces the fragility of long-horizon CAD code generation by explicitly modeling hierarchical decomposition, geometric constraints, staged generation, and progressive curriculum learning, making it a significant advancement in the field of text-to-CAD.

Overview: Claude Banana is an AI agent designed to help developers create perfect image generation prompts by translating rough ideas into detailed instructions. This tool uses a 7-component formula and 9 creative techniques to ensure the best possible results. Key features: - Asks for quick questions to understand the vision - Builds a detailed, optimized prompt using a proven formula - Provides a ready-to-copy prompt for image generators - Uses a 7-piece formula with 9 expertise modes - Knows 70+ creative techniques to enhance image quality - Works with existing prompts from Midjourney, DALL-E, or Stable Diffusion Tech stack: Python Why it matters: Claude Banana simplifies the process of creating perfect image prompts, making it easier for developers to achieve stunning visuals without needing to write detailed instructions. Repository: Hainrixz/claude-banana GitHub stars: 108 GitHub description: AI agent that turns your rough ideas into perfect image generation prompts.

Overview: JARVIS-AI-ASSISTANT is a fully autonomous AI assistant written in Python, inspired by the Iron Man movie Jarvis. It features offline speech recognition and text-to-speech capabilities using the ALICE backend, Python 3.8.9, and the vosk engine. The assistant is compatible with Windows, macOS, and Linux, and is designed for a more reliable and powerful chat mode. Key features: - Fully autonomous with AIML and Lisp as the brain of JARVIS. - Supports offline speech recognition with the vosk engine. - Offers offline text-to-speech service with pyttsx3. - Has a more reliable program launching mechanism. - Features a powerful chat mode with ALICE files. Tech stack: Python 3.8.9, ALICE backend, offline speech recognition with vosk engine, pyttsx3 text-to-speech engine. Why it matters: JARVIS-AI-ASSISTANT is a significant project for AI enthusiasts and developers, offering a comprehensive solution for creating AI assistants. Its compatibility with multiple operating systems and advanced features make it a valuable tool for creating intelligent virtual assistants.

Overview: This repository is a collection of AIGC research groups, featuring 107 GitHub stars, focusing on AI and computer vision research. Key features: - AIGC Research Groups - A Collection of AIGC Research Groups - Related Collections - Canada and China Research Groups Tech stack: Unknown Why it matters: It serves as a comprehensive resource for researchers interested in AI and computer vision, providing a platform for sharing and collaborating on research projects.

### Overview: This repository provides a PyTorch implementation of PerCo, an image compression technique that achieves perfect realism at ultra-low bitrates. It builds upon the work presented in "Towards Image Compression with Perfect Realism at Ultra-Low Bitrates" (ICLR 2024), using Stable Diffusion v2.1 as the latent diffusion model. ### Key Features: - PyTorch implementation - Uses Stable Diffusion v2.1 as the latent diffusion model - Differentiates from the official work by using the official hyper-encoder design - Includes ablation studies on LoRA and FSQ - Provides visual and quantitative performance comparisons ### Tech Stack: PyTorch ### Why it Matters: PerCo offers a significant improvement in image compression quality compared to the official work, achieving perfect realism at ultra-low bitrates. It is particularly useful for applications requiring high fidelity images, such as medical imaging and security. ### GitHub Stars: 107 ### Updates: - Finetuned whole U-Net (not just linear layers) - Slightly improved results (limited to 50k optimization steps) - Released pre-trained models - Ablation studies: experimented with LoRA and FSQ (no improvements achieved) - Switched back to official hyper-encoder design, resolved training instabilities - Significantly improved results (limited to 50k optimization steps) ### Visual Impressions: - Provides visual comparisons on the Kodak dataset for different bit-rates - Demonstrates the impact of global conditioning on the reconstructed

Overview: This project demonstrates the process of training a Convolutional Neural Network (CNN) to recognize the wake word "stop" using TensorFlow and TensorFlow Lite. It also showcases how to deploy the trained model on a Raspberry Pi for real-time speech recognition. Key features: - Training a CNN for wake word detection - Deploying the model using TensorFlow Lite for Raspberry Pi inference Tech stack: Python, TensorFlow, TensorFlow Lite, Jupyter Notebook Why it matters: This project provides a practical demonstration of how to leverage machine learning for real-world applications, specifically in the domain of wake word detection and speech recognition. It serves as an educational tool for developers looking to integrate machine learning into their projects.

### Overview: This repository is a collection of surveys, datasets, and models focused on text-driven human motion generation, particularly focusing on single-person scenarios without human-object/scene interaction. ### Key Features: - Surveys: Comprehensive reviews of generative approaches and benchmarks in motion generation. - Datasets: A curated dataset repository for motion generation tasks. - Models: Models and datasets for text-driven motion generation. ### Tech Stack: TypeScript ### Why it Matters: This repository serves as a valuable resource for researchers and developers interested in the field of text-driven motion generation. It provides a structured way to explore and contribute to the growing body of work in this area. ### GitHub Stars: 107 ### Repository: Zilize/awesome-text-to-motion ### Primary Language: TypeScript ### GitHub Description: Text-driven human motion generation surveys, datasets, and models, focusing on single-person scenarios without human-object/scene interaction.

Overview: This project presents an official implementation of the "Hierarchical Masked 3D Diffusion Model for Video Outpainting" at ACM MM 2023. It offers a solution for video outpainting, enhancing the quality of video clips by filling in missing or damaged areas. Key features: - Hierarchical Masked 3D Diffusion Model - Video Outpainting Algorithm - High-Quality Video Restoration Tech stack: Python Why it matters: This project showcases advanced techniques in video processing and restoration, providing a valuable tool for enhancing video quality and improving user experiences in applications like video editing and content creation.