Projects

Overview: LeptonAI is a Python library and command-line tool designed to simplify the operation of the NVIDIA DGX Cloud Lepton platform. Key features: - A `lep` command-line tool for creating and managing endpoints, Dynamo graph deployments, batch jobs, dev pods, Ray clusters, fine-tuning jobs, storage, secrets, and more. - A `Client` to call deployed endpoints as native Python functions. - Pythonic configuration specifications that are easily shipped to the cloud. - Skills that enable agents to operate the Lepton platform. Tech stack: Python Why it matters: LeptonAI simplifies the complex process of deploying and managing AI services on the NVIDIA DGX Cloud Lepton, making it accessible and easy to use for developers.

Overview: This project aims to use deep learning to crack captcha verification codes. It involves building a convolutional neural network using Keras to recognize captcha images. Key features: - Uses deep learning techniques for captcha recognition - Utilizes Keras for building and training the model - Supports both image and text captcha types - Includes a data generator for efficient data generation - Provides visualization code for better understanding and debugging Tech stack: TensorFlow, Keras Why it matters: This project is valuable for developers looking to automate captcha recognition, enhancing security and user experience in applications that use captcha verification. Repository: ypwhs/captcha_break Primary language: Jupyter Notebook GitHub stars: 2821 GitHub description: Crack captcha verification codes using deep learning.

### Overview: PromptBench is a unified evaluation framework designed for large language models (LLMs). It provides a comprehensive set of tools and methods to evaluate and understand the performance of various LLMs, including prompt engineering techniques and adversarial prompt evaluation. ### Key Features: - **Prompt Engineering:** Implementing efficient methods like Few-shot Chain-of-Thought, Emotion Prompt, and Expert Prompting. - **Adversarial Prompt Evaluation:** Integrating prompt attacks to assess the robustness of LLMs. - **Unified Evaluation:** Facilitating easy setup and evaluation of LLMs across different datasets and models. ### Tech Stack: - **Python:** The primary language used for development and evaluation. ### Why It Matters: PromptBench simplifies the evaluation process for researchers, making it easier to compare and understand the performance of different LLMs. Its unified approach ensures that evaluations are consistent and comparable across various datasets and models, facilitating better insights into model performance and improvement strategies. ### Repository: - **Repository:** microsoftarchive/promptbench - **GitHub Stars:** 2820 - **GitHub Description:** A unified evaluation framework for large language models. ### Installation: - **Dependencies:** Python 3.8+ and PyTorch 1.10+ are required. - **Setup:** Clone the repository and install the necessary dependencies. ### Usage: - **Quick Model Performance Assessment:** Use the provided APIs to quickly build, load datasets, and evaluate model performance. - **Prompt Engineering:** Implement prompt engineering methods to enhance model performance and understanding. ### Supported D

### Overview: TrustGraph is a Semantic Agent Governance Platform that turns enterprise knowledge and systems into an ontology-grounded operating graph connected to a policy runtime. Agents propose actions, and TrustGraph checks them against business meaning, authority, and evidence before execution, preserving the evidence and policy versions as queryable W3C PROV-O provenance. ### Key Features: - **Ontology Grounding**: Builds a connected operating graph of entities, relationships, policies, and evidence from documents and systems. - **Govern Agent Actions**: Enforces explicit rules for authority, scope, risk, uncertainty, time, and value at the point of action. - **Verify Decisions**: Retains the evidence and policy versions needed to explain and reproduce each outcome. ### Tech Stack: TrustGraph is built on W3C RDF 1.2, OWL, SHACL, and PROV-O, ensuring total transparency and portability of business knowledge. ### Why It Matters: TrustGraph provides a common semantic understanding for AI agents, enabling them to communicate effectively and make informed decisions based on the specific context and authority of the entities involved.

Overview: TeaQL Agent Kit is a model-driven harness for AI coding agents, enabling semantic modeling, app generation, progressive API Assist, and evidence-based verification. Key features: - Model-first workflow for business modeling - Evaluation and repair of business models - Generation of applications using a coding agent - Verification of application results - Integration with various programming languages and databases Tech stack: Python Why it matters: TeaQL Agent Kit provides a model-first workflow for AI coding agents, allowing developers to focus on modeling their business without needing to learn the runtime, generator, or toolchain architecture. This framework helps in building applications with a small order-management example, demonstrating how the model, generated library, runtime, and progressive Assist can be integrated effectively.

### Overview: Fractal is a high-performance distributed computing ecosystem designed to orchestrate large-scale federated learning across decentralized edge devices, ensuring user privacy and device health. ### Key Features: - **FractalCore (Central Server & Orchestrator):** Python-based control plane for multi-tenant isolation, dataset segment allocation, model graph compilation/slicing, federated weight aggregation (FedAvg), compute budgeting, and tokenized reward distribution. - **FractalAndroid (Edge Compute Node & Application):** Native Kotlin/Android execution engine running on mobile hardware, performing hardware-gated on-device training and memory-mapped inference, dynamically pausing workloads if thermals or battery thresholds are exceeded. ### Tech Stack: Python ### Why it Matters: Fractal bridges the gap between centralized model orchestration and decentralized edge intelligence, ensuring efficient and secure distributed computing while preserving user privacy and device health. Repository: Fractal-Compute-Orchestrations/FractalWorkspace GitHub Stars: 2816

### Overview: Skillshare is a versatile tool for managing and sharing AI skills, agents, rules, and more across multiple AI CLI tools. It simplifies the process of switching between different AI platforms while ensuring your skills and resources remain accessible and up-to-date. ### Key Features: - **Sync Skills and Agents**: Easily share and sync skills and agents across multiple AI tools with a single command. - **CLI and UI Integration**: Utilize the CLI for quick access and the UI for a more user-friendly experience. - **Version Control**: Keep your source code in version control, making it easy to switch tools and manage your setup. - **Team Collaboration**: Share skills and project resources with your team, allowing for easy distribution and collaboration. ### Tech Stack: Go ### Why It Matters: Skillshare simplifies the process of managing and sharing AI skills across different tools, ensuring your setup remains consistent and accessible. It allows for easy switching between platforms, version control, and seamless team collaboration, making it a valuable tool for developers and AI enthusiasts alike. ### Repository: - **Repository**: runkids/skillshare - **GitHub Stars**: 2813 - **GitHub Description**: 📚 Sync skills, agents, MCP, plugins to all AI CLI tools with one command and simplify team sharing. - **GitHub Description**: 📚 Sync skills, agents, MCP, plugins to all AI CLI tools with one command and simplify team sharing. - **GitHub Description**: 📚 Sync skills, agents, MCP, plugins to all AI CLI tools with one command and simplify team

### Overview: Kandinsky 2.2 is a multilingual text-to-image latent diffusion model that leverages improved CLIP-ViT-G for enhanced image generation and better text understanding. It includes ControlNet for precise image control, offering more accurate and visually appealing results. ### Key Features: - **Improved Text-to-Image Generation**: Enhanced by CLIP-ViT-G, leading to more aesthetically pleasing and textually coherent images. - **ControlNet Mechanism**: Enables precise image control, resulting in more accurate and visually appealing outputs. - **Latent Diffusion U-Net**: A powerful diffusion model for generating high-quality images. - **Text Encoder (XLM-Roberta-Large-Vit-L-14)**: 560M parameters. - **Diffusion Image Prior (1B)**: Maps text embeddings to image embeddings. - **CLIP Image Encoder (ViT-L/14)**: 427M parameters. - **MoVQ Encoder/Decoder (67M)**: Efficient for latent space manipulation. ### Tech Stack: The model is built using Jupyter Notebook, leveraging Python libraries for text generation and image processing. ### Why It Matters: - **Enhanced Image Quality**: CLIP-ViT-G improves the aesthetic and textual fidelity of generated images. - **Precise Control**: ControlNet allows for precise image manipulation, opening new possibilities in text-guided image generation. - **Versatile Applications**: Suitable for a wide range of applications, including fine-tuning and blending images with text. ### Repository: The

### Overview Hercules is an advanced Git repository history analysis tool written in Go. It provides fast, insightful, and highly customizable insights into a repository's commit history. Hercules is designed to be highly customizable and can be used in various organizations, including internal projects at source{d}. ### Key Features - **Fast Analysis**: Hercules is designed to be highly efficient, making it suitable for large repositories. - **Customizable Pipelines**: Users can define and execute custom analysis tasks using the plugin system. - **Built-in Analyses**: Hercules includes built-in analyses such as project burndown, files, people, churn matrix, code ownership, couples, structural hotness, aligned commit series, added vs changed lines through time, efforts through time, sentiment, and everything in a single pass. - **Plugins**: Users can extend Hercules with plugins for additional analyses. ### Tech Stack - **Go**: Written in Go, leveraging the go-git library. - **Python**: Python script for showing predefined plots. ### Why It Matters Hercules is crucial for organizations looking to gain advanced insights into their Git repository history. It helps in project management, code ownership, team collaboration, and identifying trends and patterns in the codebase. ### Installation - **Build from Source**: Hercules can be built from source using the provided Makefile. - **GitHub Action**: Hercules is also available as a GitHub Action for easy integration into CI/CD pipelines. ### Roadmap - Hercules is currently in development, with the main author back from the limbo and gradually resuming the development.

### Overview: Gerev is an AI-powered enterprise search engine designed for organizations. It enables quick access to conversations, documents, and internal pages, making it an essential tool for help desk techies and anyone looking to streamline their organization's search capabilities. ### Key Features: - **Live Online Demo:** Access a live demo of Gerev on [this link](https://app.klu.so/signup?utm_source=github_gerevai). - **Integrations:** Supports Slack, Confluence, Jira, Google Drive (.docx, .pptx), Confluence Cloud, Bookstack, Mattermost, RocketChat, Gitlab Issues, Zendesk, Stackoverflow Teams, Azure DevOps, Phabricator, Trello, Notion, Asana, Sharepoint, Box, Dropbox, and GitHub Enterprise. - **Natural Language Search:** Enables searching using natural language queries such as "How to do X," "how to connect to Y," and "Do we support Z." - **Self-Hosted Option:** Offers a self-hosted version with multi-user support, allowing organizations to host their own instance. - **GPU Machine Support:** Available on a GPU machine for enhanced performance. ### Tech Stack: Python ### Why It Matters: Gerev is designed to help organizations streamline their search capabilities, making it easier for help desk techies and anyone looking to improve their organization's efficiency. Its integration with various tools and support for natural language queries make it a powerful tool for anyone looking to enhance their search experience.

Overview: Manga OCR is an optical character recognition tool designed specifically for Japanese manga text, offering high-quality recognition of vertical, horizontal, and multi-line text, as well as text with furigana and overlays. Key features: - Custom end-to-end model using Hugging Face's Vision Encoder Decoder framework - Supports recognizing multi-line text in a single forward pass - Robust against various manga-specific scenarios Tech stack: Python Why it matters: Manga OCR is essential for manga enthusiasts and creators who need accurate text recognition for their work, ensuring consistent and high-quality results in their manga projects.

### Overview: This repository contains the code for the book "AI Agents - The Definitive Guide". It serves as a companion resource for the book, providing practical examples, tutorials, and additional material for developers. ### Key features: - Code examples for various chapters - Chapter-by-chapter notebooks - Links to open notebooks directly in Colab ### Tech stack: Jupyter Notebook ### Why it matters: The code in this repository is designed to help developers understand and implement AI agents, providing practical examples and insights from the book "AI Agents - The Definitive Guide".

### Overview: A simple artificial intelligence project that uses Neural Networks and a Genetic Algorithm to teach Google Chrome's Dinosaur to jump cacti. The project demonstrates the implementation of a Genetic Algorithm for improving neural network performance through natural selection, crossover, and mutation. ### Key features: - Neural Network implementation using Synaptic library - Genetic Algorithm for improving neural network performance - Pixel reading and keyboard simulation using RobotJs - Cross-platform compatibility (Node.js) - Cross-genome selection and mutation - Continuous learning and adaptation - Implementation of a game bug fix ### Tech stack: Node.js, Synaptic library, RobotJs, Genetic Algorithm, Pixel reading, Keyboard simulation ### Why it matters: This project showcases the integration of machine learning techniques with game development, demonstrating how AI can be applied to improve gameplay mechanics. It also highlights the importance of continuous learning and adaptation in AI systems.

### Overview: This project introduces the Keras Attention Layer, a specialized component for the Keras framework that allows for the implementation of attention mechanisms within deep learning models. The attention mechanism enhances the model's ability to focus on relevant parts of the input data, thereby improving its performance on tasks such as sequence tagging, machine translation, and more. ### Key Features: - **Attention Layer Implementation**: Provides a flexible way to incorporate attention mechanisms into Keras models. - **Score Functions**: Supports two types of attention scores: Luong and Bahdanau, enabling the model to choose the most appropriate attention mechanism based on the task requirements. - **Compatibility**: Works seamlessly with TensorFlow 2.x versions, ensuring compatibility across different TensorFlow versions. ### Tech Stack: Python ### Why it Matters: Attention mechanisms are crucial in many natural language processing (NLP) tasks, where understanding the context of input sequences is essential. By implementing this attention layer, developers can significantly enhance the performance of their models, leading to better accuracy and more efficient training processes.

### Overview: PyTorch/XLA is a Python package that leverages the XLA deep learning compiler to seamlessly integrate the PyTorch deep learning framework with Google's Cloud TPUs, enabling accelerated training and inference on Google's infrastructure. ### Key Features: - **XLA Integration**: PyTorch/XLA seamlessly integrates PyTorch with the XLA compiler, accelerating training and inference on Google's TPU devices. - **Free Access**: Try PyTorch/XLA for free on a single Cloud TPU VM with Kaggle, making it accessible for both research and production use. - **Kaggle Notebooks**: Explore examples like Stable Diffusion and Distributed PyTorch/XLA Basics using Kaggle notebooks. ### Tech Stack: PyTorch/XLA utilizes the XLA deep learning compiler to connect PyTorch and Google's Cloud TPUs. ### Why It Matters: PyTorch/XLA provides a powerful and efficient way to train deep learning models on Google's TPU devices, leveraging the optimized performance and features of XLA to accelerate the training process.

Overview: Vulnhuntr is a zero-shot vulnerability discovery tool that uses large language models (LLMs) and static code analysis to identify complex security-bypassing vulnerabilities, including RCE, SSRF, XSS, SQLI, and IDOR, in Python codebases. Key features: - Identifies locally and remotely exploitable vulnerabilities - Supports Python codebases - Automatically creates and analyzes entire code call chains - Can be run via Docker, pipx, or directly from source Tech stack: Python Why it matters: Vulnhuntr is the world's first tool to discover zero-day vulnerabilities autonomously using LLMs, making it a significant advancement in the field of cybersecurity. Repository: protectai/vulnhuntr

### Overview: This repository contains annotated research papers that I have found to be both interesting and useful. Whether you are a researcher looking to stay updated in the field of machine learning, or a developer eager to understand the intricacies of research papers, these annotated papers are designed to make the reading experience more accessible. ### Key Features: - **Adaptive Risk Minimization**: Discusses how to minimize risk in machine learning models. - **Axial DeepLab**: A paper on a novel deep learning architecture for medical image segmentation. - **ConvNext**: Explores the advancements in the field of convolutional neural networks. - **EfficientNetsV2**: A detailed look at the improvements in efficiency and performance of neural network architectures. - **Flow-edge Guided Video Completion**: A paper on techniques for video completion using flow edges. - **Is Batch Norm Unique?**: Examines the uniqueness of batch normalization in deep learning models. - **Knowledge Distillation**: A paper on the concept of knowledge distillation and its effectiveness in training neural networks. - **RandConv**: A paper on random convolutional neural networks. - **Polyloss**: A paper on the polyloss method for training neural networks. - **Scaling Down Deep Learning**: Discusses the challenges and solutions in scaling up deep learning models. - **Segment Anything**: A paper on the advancements in semantic segmentation using the Segment Anything model. - **Supervised Contrastive Learning**: A paper on the use of contrastive learning for supervised learning. - **Vision Transformer**: A detailed exploration of the vision transformer model for image processing. ###

Overview: The Unsplash Dataset is a collection of over 7.4 million high-quality Unsplash photos, keywords, and searches, designed for research and machine learning purposes. Key features: - Full dataset available for non-commercial usage - Lite dataset for commercial and non-commercial usage - Contains 7.4 million photos, 1.6 million keywords, and over 160 million searches - Offers two versions: Lite and Full Tech stack: Jupyter Notebook Why it matters: The dataset provides a rich source of data for researchers and developers to explore and discover interesting connections, enabling new research opportunities and applications.

### Overview: Clip retrieval is a tool designed to compute clip embeddings and build a clip retrieval system. It processes 100M text+image embeddings in just 20 hours using a 3080 GPU. The system includes features like a clip client for remote querying, clip inference for quick embedding computation, clip index for efficient data indexing, clip filter for data filtering, and a clip back for hosting indices with a simple Flask service. The clip front provides a user-friendly interface for querying the backend. ### Key features: - Clip client for remote querying of the backend via Python. - Clip inference for quick computation of image and text embeddings. - Clip index for building efficient indices from the embeddings. - Clip filter for filtering out data using the clip index. - Clip back for hosting indices with a simple Flask service. - Clip front for a simple UI querying the backend. - Clip end2end for a seamless end-to-end process of image to dataset, inference, index, and back and front. ### Tech stack: Jupyter Notebook ### Why it matters: Clip retrieval simplifies the process of semantic search by providing a straightforward way to compute and index clip embeddings, making it easier to build and use a semantic search system. It's particularly useful for large-scale datasets and can be applied to various applications requiring efficient image and text retrieval. ### Repository: rom1504/clip-retrieval ### Primary language: Jupyter Notebook ### GitHub stars: 2799 ### GitHub description: Easily compute clip embeddings and build a clip retrieval system with them

### Overview: YAYI 2 是中科闻歌研发的**新一代开源大语言模型**，参数规模为 30B。该模型基于 Transformer 架构，采用了超过 2 万亿 Tokens 的高质量、多语言语料进行预训练。通过百万级指令微调和人类反馈强化学习，模型能够更好地与人类价值观对齐，适用于通用和特定领域的应用场景。 ### Key features: - **Model Size**: 30B - **Pre-training**: 2 万亿 Tokens 高质量多语言语料 - **Micro-tuning**: 百万级指令微调 - **Human Feedback**: 人类反馈强化学习 - **Scoring**: C-Eval、MMLU、AGIEval、CMMLU、GAOKAO-Bench、GSM8K、MATH、BBH、HumanEval、MBPP ### Tech stack: Python ### Why it matters: 通过开源 YAYI 2 模型，我们希望促进中文预训练大模型开源社区的发展，并积极为构建雅意大模型生态做出贡献。我们希望通过这一开源项目，与每一位合作伙伴共同推动雅意大模型技术的发展，从而为更多人提供更好的语言模型服务。