Projects

### Overview: TIDY is an offline semantic Text-to-Image and Image-to-Image search application powered by a quantized state-of-the-art vision-language pretrained CLIP model and ONNX Runtime inference engine. It is designed for Android devices, ensuring offline functionality and privacy. ### Key Features: - **Text-to-Image Search**: Retrieve images based on detailed textual descriptions. - **Image-to-Image Search**: Find visually similar images by choosing from your device's gallery. - **Privacy and Security**: Operates entirely offline, ensuring no data is uploaded or shared with third parties. ### Tech Stack: Kotlin ### Why it Matters: TIDY offers a unique and powerful way to discover images based on text descriptions, providing a more intuitive and detailed search experience compared to traditional keyword-based methods. Its offline functionality ensures users can access the app without internet connectivity, enhancing accessibility and convenience.

Overview: MyVision is a free online image annotation tool designed for generating computer vision-based machine learning training data, making it easier and faster for users to label large datasets. Key features: - Drawing bounding boxes and polygons to label objects - Enriched polygon manipulation for editing, removing, and adding points - Support for various dataset formats - Pre-trained COCO-SSD model for automatic object annotation - Importing existing annotation projects and converting datasets between formats Tech stack: JavaScript Why it matters: MyVision simplifies the process of creating training data for computer vision models, making it accessible and efficient for developers and researchers. Repository: OvidijusParsiunas/myvision Primary language: JavaScript GitHub stars: 609

### Overview: Deita is an open-source project designed to facilitate automatic data selection for instruction tuning in Large Language Models (LLMs). ### Key features: - **Open-sourced Toolkits**: For automatic data selection in instruction tuning. - **Deita Datasets**: A series of extremely lightweight, high-quality alignment SFT data, released in 6k and 10k sizes. - **Deita Models**: Powerful models on par with SOTA chat LLMs, trained with 10x less instruction tuning data compared to other SOTA LLMs. ### Tech stack: Python ### Why it matters: Deita provides a solution for efficiently selecting high-quality data for instruction tuning in LLMs, potentially leading to improved performance and efficiency in model training.

### Overview: This project is a content-based movie recommendation system that uses sentiment analysis to recommend movies similar to the user's preferences. It integrates Flask for the backend, Jupyter Notebook for the frontend, and utilizes the TMDB API for fetching movie details and sentiment analysis on user reviews. ### Key Features: - Content-Based Recommendation - Sentiment Analysis on Reviews - Integration with Flask for Backend - Integration with Jupyter Notebook for Frontend - Uses TMDB API for Fetching Movie Details and Sentiment Analysis ### Tech Stack: - Flask - Jupyter Notebook - TMDB API ### Why it Matters: This system provides a personalized movie recommendation experience by leveraging the user's preferences and the sentiments expressed in their reviews. It helps users discover movies that align with their tastes and enhances the overall user experience by integrating sentiment analysis, which is crucial for understanding and improving the quality of recommendations. ### Repository: The repository is hosted on GitHub and contains the following files: - `requirements.txt`: List of required Python libraries. - `main.py`: The main script for running the application. - `static/recommend.js`: JavaScript file for recommendation logic. - `README.md`: Detailed documentation and instructions. ### GitHub Stars: The repository has been starred by 609 developers, indicating its popularity and usefulness in the community. ### Source Code: The complete source code is available at [https://github.com/kishan0725/AJAX-Movie-Recommendation-System-with-Sentiment-Analysis](https://github.com/kishan

Overview: Lite Transformer is a lightweight neural architecture designed for efficient language modeling, specifically optimized for training on limited computational resources. It features a Long-Short Range Attention mechanism to balance between speed and accuracy. Key features: - Long-Short Range Attention - Efficient memory usage - Supports both CPU and GPU training Tech stack: Python, PyTorch, fairseq Why it matters: Lite Transformer offers a significant performance improvement over traditional transformer models, making it ideal for applications requiring minimal computational resources, such as edge devices or resource-constrained environments. Its optimized architecture allows for faster training times and lower memory usage, making it a valuable tool in the field of natural language processing. Repository: mit-han-lab/lite-transformer GitHub stars: 609

Overview: This project is a PyTorch implementation of the Self-Supervised Learning of 3D Human Pose using Multi-view Geometry paper, presented at CVPR 2019. It focuses on 3D human pose estimation without requiring 3D ground-truth data or camera extrinsics. Key features: - EpipolarPose: A self-supervised learning method for 3D human pose estimation. - Training: Estimates 2D poses from multi-view images and uses epipolar geometry to obtain 3D poses and camera geometry. - Testing: Produces 3D pose results from an RGB image. - Sample outputs: Show results on the Human3.6M dataset. Tech stack: PyTorch, NVIDIA GPUs. Why it matters: This method offers a novel approach to 3D human pose estimation that does not require any ground-truth data or camera extrinsics, making it more practical and efficient.

### Overview WebShop is a simulated e-commerce website environment designed for evaluating the performance of language agents in real-world web interactions. It features 1.18 million real-world products and 12,087 crowd-sourced text instructions, simulating a complex shopping experience. The environment includes challenges such as understanding compositional instructions, query formulation, dealing with noisy text, and strategic exploration. ### Key Features - **Product Data**: 1.18 million real-world products. - **Instruction Data**: 12,087 crowd-sourced text instructions. - **Challenges**: Compositional instructions, query formulation, noise handling, and strategic exploration. ### Tech Stack Python ### Why It Matters WebShop provides a robust platform for evaluating and improving language agents in real-world web interactions, making it invaluable for researchers in natural language processing and related fields.

Overview: FlashLearn is a powerful tool that simplifies the integration of Large Language Models (LLMs) into any pipeline, providing a simple interface for orchestrating tasks and pipelines. It supports the use of various LLMs like LiteLLM, Ollama, OpenAI, and DeepSeek, and includes built-in concurrency support. Key features: - Seamlessly integrates with any existing pipeline - Supports JSON-driven flows for easy understanding and maintenance - Provides a compact JSON definition for each step/task - Built-in concurency support for efficient handling of multiple requests Tech stack: Python Why it matters: FlashLearn simplifies the process of incorporating LLMs into workflows, making it easier to integrate and maintain complex pipelines without the need for extensive code changes. The JSON-based approach ensures that the pipeline is easy to understand and maintain, while the built-in concurrency support allows for efficient handling of multiple requests, improving overall performance.

### Overview: This repository offers a collection of radar and optical satellite datasets specifically curated for ship detection, classification, semantic segmentation, and instance segmentation tasks. Ideal for applications in computer vision, machine learning, remote sensing, and maritime analysis, these datasets are highly valuable for researchers and developers working in the field. ### Key Features: - **Ship Detection**: Datasets like SAR Ship Detection Dataset (SSDD) and High-Resolution SAR Images Dataset (HRSID) are available. - **Semantic Segmentation**: Datasets such as AIR-SARShip-1.0 and LS-SSDD-v1.0 provide semantic segmentation capabilities. - **Instance Segmentation**: Datasets like FUSAR-Ship Dataset v1.0 and DSSDD offer instance segmentation features. - **Classification**: Datasets like OpenSARship-1.0 and SAR-Ship-Dataset are suitable for classification tasks. - **Regression**: No datasets provided for regression tasks. ### Tech Stack: The datasets are primarily based on radar and optical satellite imagery. ### Why It Matters: These datasets are crucial for applications in maritime surveillance, naval operations, and the development of autonomous systems, making them invaluable for researchers and developers in the field of remote sensing and computer vision. ### Repository: Repository: jasonmanesis/Satellite-Imagery-Datasets-Containing-Ships GitHub Stars: 607 GitHub Description: This repository provides a comprehensive list of radar and optical satellite datasets curated for ship detection, classification, semantic segmentation, and instance segmentation tasks. Ideal for applications in computer vision,

Overview: MIGC and MIGC++ are advanced multi-instance generation controllers designed for text-to-image synthesis and image synthesis, respectively. These models are part of the CVPR 2024 and TPAMI 2024 conference highlights. Key features: - MIGC: A controller for text-to-image synthesis. - MIGC++: An advanced controller for image synthesis that supports both instance location and instance control. Tech stack: The project uses Python as the primary language and is built on top of the Stable Diffusion models. Why it matters: These models are significant advancements in the field of text-to-image synthesis and image synthesis, offering improved control over instance locations and instance instances, respectively. They are highly relevant for researchers and developers working on AI-driven image generation and synthesis tasks.

### Overview: Radial Attention is a scalable sparse attention mechanism designed for video diffusion models. It translates spatiotemporal energy decay observed in attention score distributions into exponentially decaying compute density, achieving O(n log n) complexity while preserving expressive power for long videos. This approach enables faster inference and better performance compared to dense attention. ### Key Features: - **Physics-Inspired Sparsity**: Static masks enforce spatially local and temporally decaying attention, mirroring energy dissipation in physical systems. - **Efficient Length Extension**: Pre-trained models scale to 4× longer videos via lightweight LoRA tuning, avoiding full-model retraining. ### Tech Stack: - Python ### Why It Matters: Radial Attention accelerates pre-trained models by 1.9× at default video length and reduces tuning costs by up to 4.4×, while significantly speeding up inference by up to 3.7×. This makes it highly effective for generating high-quality videos efficiently.

### Overview: This repository provides a simplified implementation of the HRNet multi-person human pose estimation model, compatible with official pre-trained weights and supporting various models and inference options. ### Key features: - **HRNet Pose Estimation**: A simplified version of the HRNet model for multi-person pose estimation. - **YOLOv3/YOLOv3-tiny/YOLOv5 Support**: Multi-person support with YOLOv3, YOLOv3-tiny, or YOLOv5 by Ultralytics. - **TensorRT Support**: Enables inference using TensorRT for faster performance. - **Compatibility**: Fully compatible with official pre-trained weights and results. - **Python Implementation**: Written in Python, compatible with PyTorch >= 1.0. - **Training**: Includes scripts for training the HRNet network on the COCO dataset. - **Live Demo**: A reference code for running a live demo using webcam or video files. - **Jupyter Notebook**: An updated Jupyter notebook for easy demonstration and experimentation. ### Tech stack: Python ### Why it matters: This implementation provides a user-friendly interface for developers to quickly integrate multi-person pose estimation into their projects, leveraging the power of HRNet and YOLOv3/YOLOv5 for efficient inference and compatibility with various models.

Overview: 本项目专注于法律裁判文书的事件抽取及其应用，包括数据预处理、分词、词性标注、命名实体识别、事件要素抽取和判决结果预测等内容。 Key features: - 数据预处理 - 分词、词性标注、命名实体识别 - 基于CRF的事件要素抽取 - 判决结果预测 - 案件相似度分析 Tech stack: Python Why it matters: 本项目通过法律裁判文书的事件抽取，旨在为法律文本处理提供有效工具，提高案件相似度分析的准确性和判决结果预测的精度，从而提升法律服务的效率和质量。

### Overview: Jev is a powerful tool for software development, providing a typed judgment to ensure code remains in control. It offers a comprehensive guide to TypeSafe's System One model, including SDKs, live demos, agent tools, and independent evaluations. ### Key features: - **Documentation**: Detailed documentation on how to use Jev. - **Live Projects**: Access to live projects for hands-on experience. - **Starter Kits**: Pre-configured Jev setups for quick start. - **Independent Evaluations**: Verified by independent developers and experts. ### Tech stack: HTML ### Why it matters: Jev simplifies complex decision-making processes in software development, ensuring that code remains focused and controlled. It provides a structured approach to handling typed questions and decisions, making it an invaluable resource for developers looking to improve their coding practices. ### GitHub stars: 606 ### Repository: AbdelStark/awesome-typesafe-jev ### Primary language: HTML ### GitHub description: A community field guide to TypeSafe's Jev: see one documented call, try live projects, copy a starter, and inspect independent tests. ### Repository: AbdelStark/awesome-typesafe-jev ### Starred on GitHub: 606

### Overview: This project presents a new algorithm for converting binary images into a set of polylines, known as skeleton tracing. The algorithm is designed to efficiently trace the topological skeleton of an image, providing a vector representation that is more desirable than enclosing outlines. ### Key features: - **Parallelizable Divide-and-Conquer Algorithm**: The algorithm divides the image into smaller submatrices and recursively processes them, ensuring parallel execution. - **Real-time Performance**: The algorithm is designed to trace the skeleton in real time, making it suitable for applications requiring immediate feedback. - **Efficient Polylines Representation**: The output is a set of polylines that represent the skeleton, providing a clear and concise representation of the image's structure. ### Tech stack: - **Primary Language**: C - **Support Languages**: C++, Java, JavaScript, Python, Go, C#, Unity, Swift, Rust, Julia, WebAssembly, Haxe, Processing, OpenFrameworks ### Why it matters: This algorithm is particularly useful in applications where a vector representation of the skeleton is required, such as in medical imaging, computer vision, and other fields where precise and efficient representation of image structures is essential. The ability to trace the skeleton in real time and convert it into polylines makes it highly valuable for applications that require immediate feedback or where the quality of the representation is critical.

Overview: This repository contains a series of Jupyter Notebook-based notebooks designed for practicing engineers. It covers various aspects of Natural Language Processing (NLP) using Python and deep learning techniques, including text classification, spell correction, named entity recognition, text representations, modern text classification, deep learning for NLP, and building a chatbot. Key features: - Text Classification - Spell Correction - Named Entity Recognition - Text Representations - Modern Text Classification - Deep Learning for NLP - Building a Chatbot Tech stack: Jupyter Notebook Why it matters: This work builds on existing NLP textbooks and provides a practical guide for practitioners, focusing on code-first approaches and practical applications.

Overview: This project is a fullstack example of how to build a RAG (Retrieval Augmented Generation) app with Cloudflare, utilizing Cloudflare Workers, Pages, D1, KV, R2, AI Gateway, and Workers AI. It showcases the integration of various cloudflare services to create a comprehensive RAG application. Key features: - Streamed interactions using Server-Sent Events - Hybrid RAG using Full-Text Search on D1 and Vector Search on Vectorize - Switchable providers (OpenAI, Groq, Anthropic) with fallbacks using AI Gateway - Per-IP rate limiting using Cloudflare's KV - OCR running inside a Cloudflare Worker using unpdf - Smart placement for optimal application performance Tech stack: TypeScript Why it matters: This project demonstrates the capabilities of Cloudflare's services in creating a robust and efficient RAG application, showcasing the integration of various cloudflare services to achieve a comprehensive solution. Repository: RafalWilinski/cloudflare-rag GitHub stars: 607 GitHub description: Fullstack "Chat with your PDFs" RAG (Retrieval Augmented Generation) app built fully on Cloudflare Deployment: Deploy to Cloudflare Pages using the provided secrets and deployment script.

### Overview: Adaline Gateway is the only fully local, production-grade Super SDK that provides a simple, unified, and powerful interface for calling over 300+ LLMs. ### Key Features: - 🔧 Strongly typed in TypeScript - 📦 Isomorphic - works everywhere - 🔒 100% local, private, and _NOT_ a proxy - 🛠️ Tool calling support across all compatible LLMs - 📊 Batching for all requests with custom queue support - 🔄 Automatic retries with exponential backoff - ⏳ Caching with custom cache plug-in support - 📞 Callbacks for comprehensive instrumentation and hooks - 🔍 OpenTelemetry integration for existing infrastructure - 🔌 Plug-and-play custom providers for local and custom models ### Tech Stack: TypeScript ### Why it Matters: Adaline Gateway is trusted by enterprises and provides a robust, local, and powerful interface for calling over 300+ LLMs. It supports various providers and is highly extensible, making it a valuable tool for developers and enterprises alike.

Overview: AS-One is a Python library that simplifies the integration of various computer vision algorithms, including YOLO, SAM, ByteTrack, and DeepSORT, into a single modular framework. This makes it easier to implement and use these advanced computer vision techniques without needing to write extensive code. Key features: - Integration of multiple computer vision algorithms (YOLO, SAM, ByteTrack, DeepSORT) - Support for different versions of YOLO - Compatibility with ONNX, PyTorch, and CoreML flavors - Easy installation via pip Tech stack: Python Why it matters: AS-One simplifies the process of integrating complex computer vision algorithms, making it accessible and efficient for developers looking to enhance their projects with advanced vision techniques. The library's modular design allows for easy updates and improvements in the future, ensuring continued relevance and functionality.

### Overview: This repository contains material for RAW image processing, RAW Restoration and Super-Resolution, RAW reconstruction from sRGB, learned Image Signal Processing (ISP), Image Enhancement and Restoration (denoising, deblurring), Bokeh rendering, and much more! ### Key features: - RAW Restoration - RAW Super-Resolution - RAW Reconstruction from sRGB - Learned Image Signal Processing (ISP) - Image Enhancement and Restoration (denoising, deblurring) - Bokeh rendering - Many more works and challenges for the NTIRE (CVPR) and AIM (ICCV/ECCV) Challenges. ### Tech stack: Jupyter Notebook ### Why it matters: This repository provides comprehensive material for RAW image processing, RAW Restoration, and Super-Resolution, making it an essential resource for researchers and developers in the field of computer vision and imaging.