Projects

Overview: Avalanche is an end-to-end library for continual learning based on PyTorch, designed to facilitate fast prototyping, training, and reproducible evaluation of continual learning algorithms. Key features: - A uniform API for data handling - Comprehensive utilities for model training and evaluation - Pre-implemented continual learning baselines and state-of-the-art algorithms - Advanced logging and plotting features for real-time experiment tracking Tech stack: Python Why it matters: Avalanche provides a shared and collaborative open-source codebase for continuous learning research, enabling researchers to quickly prototype, train, and evaluate continual learning algorithms without the need to write extensive code.

### Overview: This repository is a collection of articles and papers on diffusion models applied in medical imaging. It aims to provide a comprehensive overview of the latest research in the field, focusing on applications such as anomaly detection, denoising, segmentation, image-to-image translation, reconstruction, image generation, and more. ### Key Features: - **Survey Papers**: Comprehensive reviews of diffusion models in medical imaging, including foundational works and recent advancements. - **Challenge Reports**: Insights into challenges and solutions in medical imaging using diffusion models. - **Papers**: A diverse range of papers covering various aspects of diffusion models in medical imaging, including anomaly detection, denoising, segmentation, image-to-image translation, and more. - **Applications**: The repository covers applications such as image restoration, editing, adversarial attacks, fairness, time series analysis, and audio processing. ### Tech Stack: The repository uses Markdown, GitHub, and various LaTeX packages for its documentation and presentation. ### Why It Matters: This collection is valuable for researchers, developers, and practitioners in the field of medical imaging, offering a platform to stay updated with the latest research and applications of diffusion models in medical imaging.

Overview: Agentset is an open-source platform designed for building, evaluating, and deploying production-ready RAG and agentic applications. It provides a complete ecosystem from ingestion, vector indexing, evaluation, and hosting. Key features: - Turnkey RAG: Ingestion, chunking, embeddings, and retrieval. - Model agnostic: Works with any LLM, embeddings, and vector DB. - Chat playground with message editing and citations. - Production hosting with preview links and custom domains. - API + typed SDKs, OpenAPI spec. - Built-in multi-tenancy. - Built with TypeScript, Next.js, AI SDK, Prisma, Supabase, and Trigger.dev. Tech stack: TypeScript, Next.js, AI SDK, Prisma, Supabase, and Trigger.dev. Why it matters: Agentset simplifies the process of building, evaluating, and deploying RAG and agentic applications, making it easier for developers to create high-performance, production-ready solutions.

### Overview: Solid Glass is a design system for AI applications, featuring a refined glass-inspired aesthetic that balances clarity and stability. It offers a balanced blend of flat design, liquid glass, and physical elements, providing a clean, precise, and tangible interface. ### Key Features: - **Flat Design**: Interfaces feel tangible with subtle touch and response. - **Liquid Glass**: Nearly clear, emphasizing precision and technical aesthetics. - **Physical Elements**: Depth comes from thickness, shadow, and spatial hierarchy, giving surfaces a physical presence. - **Materiality**: Subtle edge highlights, fine borders, and soft shadows define each surface. ### Tech Stack: TypeScript ### Why It Matters: Solid Glass is designed to be a calm and readable interface for AI applications, reducing visual distractions and focusing on meaningful feedback. It balances flat design, liquid glass, and physical elements to create a clean, precise, and tangible interface that is easy to read and navigate. ### Repository: **deeplethe/solid-glass** ### GitHub Stars: 2112 ### GitHub Description: A design system for AI applications. Material, physical, tactile, technical.

Overview: Inkloom is developing specialized AI models for logo design, transforming a business description into concrete constraints, selecting and fitting letterforms, composing geometric primitives, and producing a mark that can be described, defended, and reproduced. Key features: - Brand analysis - Typography - Symbol construction - Composition Tech stack: TypeScript Why it matters: Inkloom's models provide a studio-like approach to logo design, ensuring the mark is constructed rather than sampled, making it more versatile and easier to reproduce. Repository: Inkloom-art/inkloom Primary language: TypeScript GitHub stars: 2111 GitHub description: Specialised AI models for logo design — a brand-analysis model turns a business into constraints, typography and symbol models construct the mark, and a composition engine produces real lockups and clear-space rules. Early access open.

### Overview: Open Agent Email is an open-source email infrastructure designed for AI agents, providing secure inboxes, REST and MCP APIs, SMTP, human approval workflows, signed webhooks, and Python and TypeScript SDKs. It allows agents to receive, reply, and send emails, and includes self-hosting options with Docker and PostgreSQL. ### Key Features: - **Secure Inboxes**: Each agent has its own inbox with real addresses and conversation history. - **Human Approval Workflow**: Hold messages for approval with approve and reject links, and an expiry time. - **Self-Hosted**: Self-hosted with Docker and PostgreSQL, allowing full control over the infrastructure. - **REST and MCP APIs**: Accessible through REST, webhooks, live streams, and MCP for coding agents. - **Python and TypeScript SDKs**: SDKs for Python and TypeScript to integrate with the email infrastructure. ### Tech Stack: - Python 3.11 or newer - Docker - PostgreSQL ### Why It Matters: - **Identity and Control**: Provides a real address and mailbox for each agent, ensuring no identity leakage. - **Safety Net**: Holds messages for approval, preventing mistakes from reaching customers. - **Self-Hosted**: Allows full control over the infrastructure, with options for self-hosting and self-managing. - **APIs and SDKs**: Accessible through REST, webhooks, live streams, and MCP, making it easy to integrate with other systems. ### Repository: - **GitHub**: Anil-matcha/open-agent-email - **Stars**: 2,10

Overview: SRU (Simple Recurrent Units) is a fast recurrent neural network unit that achieves significant speed improvements over traditional LSTM units, particularly on CPU. It is designed to be as fast as word-level convolution using convolutional operations, making it highly suitable for applications where computational resources are limited. Key features: - SRU is designed to run 10 to 16 times faster than LSTM, without sacrificing accuracy. - It operates as fast as (or faster than) word-level convolution using convolutional operations. - SRU++ is a variant that offers further speed improvements and better performance on certain tasks. Tech stack: Python Why it matters: SRU is a highly efficient recurrent neural network unit that can significantly reduce the computational cost of training natural language processing models, making it ideal for applications with limited computational resources.

### Overview: ATLAS is a local coding agent designed to bring frontier-style reasoning and verification to compact open models. It enhances the intelligence around the model, enabling smaller models to handle real software work entirely on your own hardware, without the need for hosted APIs or token fees. ### Key Features: - **Adaptive Test-time Learning and Autonomous Specialization (ATLAS)** - **Model Agnostic** - **Version 3.1.6** - **License: AGPL-3.0** - **Tests: ✓** - **Install Matrix: ✓** - **CodeQL: ✓** - **Container Scan: ✓** - **Release Signature: ✓** - **OpenSSF Scorecard: ✓** ### Tech Stack: Python ### Why It Matters: - **Get more from smaller models.** ATLAS adds planning, candidate selection, verification, and repair around the model, enhancing the performance of smaller models. - **Verify before accepting.** Generated code can be compiled, tested, and corrected inside an isolated execution environment, ensuring higher quality. - **Spend compute where it matters.** Smaller edits take a shorter path, while harder tasks receive more candidates, reasoning, and validation. - **Run your own model.** Use a compatible GGUF model on NVIDIA, AMD, Apple Silicon, Vulkan, or CPU-supported hardware. - **Keep control local.** ATLAS does not upload your repository or prompts to a hosted model or ATLAS-operated service, ensuring full control over your project. - **Own the full stack.** ATLAS is open

Overview: This project implements the Convolutional Recurrent Neural Network (CRNN), a state-of-the-art model for image-based sequence recognition tasks such as scene text recognition and Optical Character Recognition (OCR). Key features: - CRNN model for image-based sequence recognition - Support for both lexicon-free and lexicon-based recognition - Pretrained model available for download and usage - Docker support for easy deployment Tech stack: Lua Why it matters: The CRNN model offers advanced capabilities for recognizing text in images, making it highly relevant in applications like scene text recognition and OCR. Its ability to handle both lexicon-free and lexicon-based recognition tasks makes it versatile and effective for various recognition scenarios.

### Overview: DevDocs is a completely free, private, UI-based Tech Documentation MCP server designed specifically for coders and software developers. It seamlessly integrates with popular tools like Cursor, Windsurf, Cline, Roo Code, and Claude Desktop App, making it easy to integrate into your workflow. ### Key Features: - **Smart Depth Control**: Choose crawl depth from 1-5 levels. - **Automatic Link Discovery**: Finds and categorizes all related content. - **Selective Crawling**: Pick exactly what you want to extract. - **Child URL Detection**: Automatically discovers and maps website structure. - **Parallel Processing**: Crawl multiple pages simultaneously. - **Smart Caching**: Never waste time on duplicate content. - **Lazy Loading Support**: Handles modern web apps effortlessly. - **Rate Limiting**: Respectful crawling that won't overload servers. ### Tech Stack: TypeScript ### Why It Matters: DevDocs streamlines the process of understanding and implementing complex technologies, reducing the time typically spent on documentation research and development. By providing a user-friendly interface and intelligent crawling capabilities, it makes it easier for developers to get up to speed quickly and efficiently.

### Overview: Humanoid-Gym is an open-source reinforcement learning framework designed for training humanoid robots, particularly focusing on zero-shot transfer from simulation to the real-world environment. It is built on Nvidia Isaac Gym and includes a sim-to-sim framework that allows users to verify policies across different simulations. ### Key Features: - **Comprehensive Training Guidelines**: Detailed walkthroughs for each stage of training. - **Step-by-Step Configuration Instructions**: Clear and concise instructions for efficient setup. - **Execution Scripts**: Pre-prepared scripts to streamline the training workflow. - **Sim2Sim Support**: Allows transferring trained policies to highly accurate simulated environments. - **Denoising World Model Learning**: Integrates state estimation and system identification for practical and effective real-world learning. ### Tech Stack: Python ### Why It Matters: Humanoid-Gym simplifies the process of training humanoid robots in real-world simulations, ensuring zero-shot transfer from simulation to the real-world environment. This framework is particularly useful for researchers and developers looking to improve the performance and robustness of RL-trained policies in real-world applications.

### Overview: Chain-of-ThoughtsPapers is a project that explores the development and application of large language models (LLMs) in generating and explaining logical reasoning. This repository aims to provide insights into the latest advancements in prompting LLMs to elicit reasoning, including self-consistency, self-teaching, and other techniques that enhance the models' ability to reason logically. ### Key Features: - **Chain of Thought Prompting**: A method to elicit reasoning from LLMs by providing context and guiding prompts. - **Self-Consistency Improves Reasoning**: Techniques that enhance the consistency of reasoning within LLMs. - **STaR: Self-Taught Reasoner**: A bootstrapping approach that improves reasoning through reasoning. - **PaLM: Scaling Language Modeling**: A pathway-based scaling method for language models. - **Can Language Models Learn from Explanations**: Research on how LLMs can learn from contextual explanations. - **Inferring Implicit Relations**: Methods for inferring implicit relations using language models. - **The Unreliability of Explanations**: Studies on the reliability of explanations in few-shot in-context learning. - **Large Language Models as Reasoners**: Research on LLMs being zero-shot reasoners. - **Least-to-Most Prompting**: Techniques that enable complex reasoning in LLMs. - **Selection-Inference**: Exploiting LLMs for interpretable logical reasoning. - **On the Advance of Making Language Models Better Reasoners**: A paper discussing advancements in making LLMs better reasoners. ### Tech Stack

### Overview: Ecco is a Python library designed to help explain and visualize Natural Language Processing (NLP) models, particularly those based on the Transformer architecture. It provides interactive visualizations directly within Jupyter notebooks, making it easier to understand and analyze the behavior of these models. ### Key Features: - **Support for Multiple Models:** Ecco supports a wide variety of language models, including GPT2, BERT, RoBERTa, T5, and T0, as well as other Transformer-based models. - **Local Model Support:** Users can add their own local models based on Hugging Face's PyTorch models. - **Feature Attribution:** Ecco includes feature attribution techniques such as IntegratedGradients, Saliency, InputXGradient, DeepLift, DeepLiftShap, GuidedBackprop, GuidedGradCam, Deconvolution, and LRP. - **Neuron Activation Capture:** It captures neuron activations in the FFNN layer of the Transformer block, enabling the visualization of neuron activation patterns. - **Neuron Activation Visualizations:** Ecco provides visualizations for the evolution of token processing, candidate output tokens, and their probabilities at each layer of the model. ### Tech Stack: - **Python:** Built on top of PyTorch and Transformers. - **Jupyter Notebook:** The primary interface for exploring and explaining NLP models. ### Why It Matters: Ecco is an essential tool for anyone looking to better understand and explain the behavior of Transformer-based NLP models. It provides a comprehensive set of visualizations and attribution techniques that can

### Overview: VisoMaster is a powerful and easy-to-use video face swapping and editing software that utilizes AI to produce natural-looking results with minimal effort. It is ideal for both casual users and professionals. ### Key Features: - **Face Swap**: Supports multiple face swapper models, including DeepFaceLab trained models (DFM), and advanced multi-face swapping with masking options for each facial part. Occlusion masking support (DFL XSeg Masking) is also available. - **Face Editor (LivePortrait Models)**: Allows manual adjustment of expressions and poses for different face parts, with fine-tuning of colors using RGB adjustments. - **Live Playback**: Real-time preview of processed video before saving. - **Face Embeddings**: Uses multiple source faces for better accuracy and similarity. - **Live Swapping via Webcam**: Stream to virtual camera for applications like Twitch, YouTube, and Zoom. - **User-Friendly Interface**: Intuitive and easy to use. - **TensorRT Support**: Leverages supported GPUs for ultra-fast processing. - **Many More Advanced Features**: Including video markers, adjustable settings per frame, and support for video markers. ### Tech Stack: Python ### Why It Matters: VisoMaster is a versatile tool that simplifies the process of face swapping and editing in images and videos, making it accessible to both casual users and professionals. Its AI-driven capabilities ensure natural-looking results with minimal effort, making it an essential tool in the field of video editing and face swapping.

Overview: This repository is dedicated to tracking the latest advancements in federated learning, including research papers, books, codes, tutorials, and videos. Key features: - Research Papers - Books - Codes - Tutorials - Videos Tech stack: Python, TensorFlow, PyTorch, and other relevant machine learning libraries. Why it matters: Federated learning is crucial for maintaining data privacy and security in distributed machine learning environments, making it an essential topic in the field of artificial intelligence and data mining. Repository: innovation-cat/Awesome-Federated-Machine-Learning GitHub stars: 2103 GitHub description: Everything about federated learning, including research papers, books, codes, tutorials, and videos.

### Overview: EfficientNet is a lightweight convolutional neural network architecture that achieves state-of-the-art accuracy on ImageNet and other commonly used transfer learning datasets. This reimplementation is compatible with both Keras and TensorFlow Keras frameworks, making it accessible for developers. ### Key Features: - **EfficientNet-B0**: The smallest EfficientNet model with 66M parameters and 37B FLOPS. - **EfficientNet-B1**: 7.6x smaller and 5.7x faster on CPU inference compared to ResNet-152. - **EfficientNet-B4**: 82.6% improvement in top-1 accuracy over ResNet-50. - **EfficientNet-B7**: 84.4% top-1 accuracy on ImageNet with 66M parameters and 37B FLOPS. ### Tech Stack: - Python ### Why It Matters: EfficientNet models are highly efficient, achieving high accuracy with minimal computational resources, making them ideal for resource-constrained environments.

### Overview: ViZDoom is a powerful reinforcement learning environment designed for developing AI bots that play the classic 1993 game Doom using only visual information. It offers a unique and immersive experience by allowing researchers to train agents to play the game purely through visual inputs, making it highly relevant for machine learning and visual reinforcement learning applications. ### Key Features: - **API for Python and C++**: Provides both Python and C++ APIs, making it accessible to developers and researchers. - **Multi-platform Support**: Runs on Linux, macOS, and Windows, ensuring compatibility across different operating systems. - **High Performance**: Capable of handling up to 7000 frames per second in sync mode, with single-threaded performance on modern CPUs. - **Lightweight**: Minimal resource usage, with a few MBs of memory footprint. - **Custom Scenarios**: Offers powerful visual editors, scripting language, and examples for easy creation of custom scenarios. - **Async and Sync Modes**: Supports both async and sync single-player and multiplayer modes. - **Customizable Settings**: Allows adjusting resolution, rendering parameters, and more. - **Access to Buffers**: Provides access to the depth buffer, audio buffer, and in-game text messages. - **Off-screen Rendering**: Enables off-screen rendering for better performance. - **Recording Episodes**: Records episodes for analysis and evaluation. - **In-game Time Scaling**: Works in both sync and async modes, offering flexibility. ### Tech Stack: - **Python**: The primary language for the API, providing a friendly environment for reinforcement learning applications.

Overview: This project is a Keras model designed to identify various types of inappropriate content, including drawings, hentai, porn, and more, from images. Key features: - The model is trained on 60+ gigabytes of data to accurately classify images into safe for work, hentai, porn, and other categories. - It powers the NSFW JS library, which can be used for programmatic detection. - The model achieves 93% accuracy based on Inception V3. Tech stack: Python Why it matters: This model is crucial for developers looking to create applications that can identify and prevent inappropriate content, ensuring a safer environment for users. Repository: GantMan/nsfw_model Primary language: Python GitHub stars: 2100 GitHub description: Keras model of NSFW detector README excerpt: 

### Overview: Zero123++ is a Single Image to Consistent Multi-view Diffusion Base Model that achieves state-of-the-art results in 3D generation and multi-view synthesis. It is designed to handle a wide range of input field of views and provides improved accuracy and robustness. ### Key Features: - **Multi-View Synthesis:** Zero123++ generates consistent multi-view images from a single input image, providing a more realistic representation of the object. - **Improved Robustness:** The model is more robust to variations in input field of views and croppings, ensuring consistent results across different scenarios. - **3D Generation:** The model focuses on generating 3D views, providing accurate and detailed representations of objects. - **Normal Generator:** A normal generator ControlNet is included, which can generate view-space normal images, enhancing the accuracy of alpha masks. - **Validation Metrics:** On the validation set from Objaverse, the model achieves an alpha (before matting) IoU of 98.81%, mean normal angular error of 10.75°, and normal PSNR of 26.93 dB. ### Tech Stack: Python ### Why It Matters: Zero123++ is essential for applications requiring high-quality 3D representation and multi-view synthesis, making it a valuable tool for developers working in fields such as computer vision, augmented reality, and virtual reality.

Overview: The Open Graph Benchmark (OGB) is a comprehensive collection of benchmark datasets, data loaders, and evaluators designed for graph machine learning. It covers a wide range of tasks and diverse datasets, including small-scale, medium-scale, and large-scale graphs, as well as datasets from various domains such as scientific, social, and heterogeneous knowledge graphs. Key features: - Easy-to-use PyTorch Geometric and DGL data loaders for automatic dataset downloading, standardized dataset splits, and unified performance evaluation. - Standardized evaluators for easy evaluation and comparison of different methods. Tech stack: Python, PyTorch, DGL, Numpy, pandas, urllib3, scikit-learn, outdated. Why it matters: OGB provides a robust and versatile platform for benchmarking and evaluating graph machine learning algorithms, ensuring that researchers can focus on developing new techniques without worrying about the complexities of dataset preparation and evaluation.