Projects

### Overview: geo-sleuth is an agent skill that locates the origin of a photo using OpenStreetMap geometry, elevation data, satellite imagery, and street view. It provides precise results with minimal human intervention, making it an invaluable tool for geolocation and evidence analysis. ### Key Features: - **OpenStreetMap Geometry**: Determines the exact location based on OpenStreetMap data. - **Elevation Data**: Uses elevation information to refine the search area. - **Satellite Imagery**: Provides high-resolution satellite images for accurate positioning. - **Street View**: Utilizes street view data for additional context and verification. ### Tech Stack: Python ### Why It Matters: geo-sleuth simplifies the process of geolocation by leveraging modern mapping technologies. It eliminates the need for extensive human intervention, making it highly efficient and reliable. The skill works seamlessly with various agent skills, enhancing their capabilities in real-world applications.

### Overview: This repository contains the TensorFlow implementation of YOLOv3, a popular object detection model. The code is designed to be compatible with TensorFlow 2.0 and includes a detailed guide on how to set up and use the model, as well as a comprehensive tutorial on training and evaluating the model on the VOC dataset. ### Key Features: - **YOLOv3 Implementation**: The repository provides a complete implementation of YOLOv3 using TensorFlow 2.0. - **Dependencies**: Users need to install some dependencies before getting started, including TensorFlow, Keras, and other necessary libraries. - **Data Preparation**: The repository includes a dataset preparation script and a class list for the VOC dataset. - **Training and Evaluation**: Detailed instructions on training the model from scratch and from pre-trained COCO weights are provided, along with evaluation metrics such as mAP on the VOC2012 dataset. ### Tech Stack: - Python - TensorFlow 2.0 ### Why it Matters: This implementation is valuable for developers looking to integrate YOLOv3 into their projects or for those interested in understanding how to build and evaluate object detection models from scratch using TensorFlow. The repository provides a practical guide that can be easily adapted for various datasets and configurations. ### Repository: - **GitHub Repository**: [YunYang1994/tensorflow-yolov3](https://github.com/YunYang1994/tensorflow-yolov3) - **Stars**: 3612 - **Description**: "🔥 TensorFlow Code for technical

### Overview: TorchRL is a modular and PyTorch-native toolkit for reinforcement learning, decision making, robotics, and simulation. It is designed to be a collection of composable pieces that can be easily swapped and scaled from a local prototype to vectorized, multiprocess, distributed, and model-based workflows without changing the data model. ### Key features: - **TensorDict Data Model**: A dictionary-like tensor container with PyTorch operations, device transfers, shared-memory support, memmaps, lazy views, and `nn.Module` wrappers. - **Independent Modules**: Environments, policies, replay buffers, objectives, and collectors are independent modules that can be swapped without rewriting the rest of the stack. - **Research Code Scaling**: Research code can scale from a local prototype to vectorized, multiprocess, distributed, compiled, recurrent, multi-agent, model-based, or offline workflows without changing the data model. ### Tech stack: PyTorch ### Why it matters: TorchRL is essential for anyone working on reinforcement learning, robotics, or simulation who wants a modular, PyTorch-native toolkit that can be easily scaled and adapted to various workflows without changing the core data model. ### Repository: pytorch/rl ### GitHub stars: 3606 ### Primary language: Python ### GitHub description: TorchRL is a PyTorch-native toolkit for reinforcement learning, decision making, robotics, and simulation. It is not a single algorithm implementation or a narrow benchmark suite: it is a collection of composable pieces for building RL systems while keeping the code close to the Py

### Overview: TruLens is a comprehensive evaluation and tracking system designed to help AI teams improve the quality and efficiency of their models. It uses LLM judges to score every step of the development process, providing detailed feedback and helping teams identify areas for improvement. ### Key Features: - **Metrics Evaluation**: TruLens scores every step of the development process, providing detailed feedback. - **RAG Triad Benchmarking**: TruLens compares its performance against established benchmarks, including a fine-tuned proprietary model. - **Contextual Relevance**: TruLens excels in context relevance, scoring the correct passage over a near-copy with one fact swapped. - **Agent Errors Detection**: TruLens accurately detects agent errors, achieving a high accuracy rate in agent GPA and other metrics. ### Tech Stack: TruLens is built using Python and follows the OpenTelemetry-native tracing standard, ensuring compatibility with various backend systems. ### Why It Matters: TruLens helps AI teams identify and address issues in their models, leading to better quality and efficiency. Its metrics and benchmarks make it a valuable tool for continuous improvement and model optimization. ### Repository: truera/trulens - **Primary Language**: Python - **GitHub Stars**: 3605 - **GitHub Description**: Evaluation and Tracking for LLM Experiments and AI Agents - **GitHub License**: MIT License - **GitHub Issues**: 100+ open issues - **GitHub Pull Requests**: 50+ open pull requests - **GitHub Description**: Evaluation and Tracking for LLM Ex

Overview: This project presents a novel Thin-Plate Spline Motion Model for image animation, which was presented at CVPR 2022. The model is designed to enhance image animations by using Thin-Plate Spline interpolation techniques. Key features: - Thin-Plate Spline Motion Model - Integration with Hugging Face Spaces for web demo - Pre-trained models available on various cloud platforms - Support for multiple datasets including VoxCeleb, TaiChiHD, and TED-talks Tech stack: Jupyter Notebook Why it matters: This model offers a significant improvement in image animation quality by leveraging advanced interpolation techniques, making it a valuable tool for enhancing visual content. Repository: yoyo-nb/Thin-Plate-Spline-Motion-Model Primary language: Jupyter Notebook GitHub stars: 3605 GitHub description: Source code of the CVPR'2022 paper "Thin-Plate Spline Motion Model for Image Animation" | Paper: https://arxiv.org/abs/2203.14367 | Supp: https://cloud.tsinghua.edu.cn/f/f7b8573bb5b04583949f/?dl=1

Overview: Apache SINGA is a distributed deep learning platform designed for high-performance training of large-scale neural networks. Key features: - Distributed training capabilities - Support for various deep learning frameworks - Scalable architecture for handling large datasets - Extensive documentation and community resources Tech stack: C++ Why it matters: Apache SINGA provides a robust and efficient solution for distributed deep learning, making it ideal for researchers and developers working on large-scale machine learning projects.

### Overview: Deta Surf is a personal AI notebook designed to streamline research and thinking by integrating files and web pages directly into a stream of thought. It uses open formats and local storage to minimize manual work, making it accessible and open-source. ### Key features: - **Multi-Media Library & Notebooks**: Store and organize files, web links, and local files in a private library that can be used offline. - **Smart Notes**: Automatically generate notes from tabs, websites, and resources in your library, including web searches, citations, and interactive applications. - **Tabs, Split View, & Sidebar**: Use tabs, split view, and a sidebar for easy navigation and quick access to your library and notes. - **Surflets (App Generation)**: Generate interactive applets to visualize, understand, or explore concepts or data. - **AI**: Powered by large language models of your choice, integrating with smart notes and surflets for enhanced functionality. ### Tech stack: Deta Surf is built using Svelte, TypeScript, and Rust, with a focus on local storage and open data formats. ### Why it matters: Deta Surf aims to help people think better by providing a personal and open platform for research and synthesis of ideas, minimizing the need for manual work and integrating seamlessly with open formats and AI models.

Overview: face.evoLVe is a high-performance face recognition library built on PaddlePaddle and PyTorch, designed to enhance comprehensiveness, effectiveness, and efficiency for face-related analytics and applications. Key features: - Integrates PaddlePaddle and PyTorch for comprehensive face recognition capabilities. - Supports training and fine-tuning of models using the PaddlePaddle framework. - Provides publicly available datasets for face anti-spoofing/liveness detection and face recognition analytics. - Offers feature extraction APIs for pre-trained models using PyTorch and OpenCV. - Supports distributed training with multi-GPUs for large-scale face recognition tasks. Tech stack: PaddlePaddle, PyTorch, Python. Why it matters: This library is crucial for researchers and developers looking to improve the performance and efficiency of face recognition systems, particularly in large-scale applications and real-world scenarios.

### Overview: FlashRAG is a comprehensive Python toolkit designed for the efficient research and development of Retrieval Augmented Generation (RAG) techniques. It includes 36 pre-processed benchmark datasets and 23 advanced RAG algorithms, covering reasoning-based methods that enhance performance on complex tasks. ### Key Features: - Extensive and customizable framework for assembling complex RAG pipelines. - Comprehensive benchmark datasets for testing and validating RAG models. - Pre-implemented advanced RAG algorithms with reported results. - Reasoning-based methods that combine reasoning ability with retrieval, achieving superior performance on complex multi-hop tasks. - Efficient preprocessing stage for simplified workflow preparation. - Optimized execution with tools like vLLM and FastChat for enhanced library efficiency. ### Tech Stack: Python ### Why It Matters: FlashRAG provides a powerful and flexible toolkit for researchers and developers working in the RAG domain, enabling the reproducibility and customization of state-of-the-art RAG algorithms.

### Overview: Sprix SAGE Router is an open-source research project aimed at state-aware routing for A2A agent networks, designed to dynamically decide which agent should execute a task based on the completed DAG nodes, reusable artifacts, and partial quality, among other factors. ### Key Features: - **State-Aware Graph Exchange (SAGE)**: Evaluates three routes (SELF, COLLABORATE, HANDOFF) to determine the best feasible agent configuration for executing tasks. - **Checkpoint-Aware Rerouting**: Preserves completed requirements and combines in-flight completion, observed partial quality, current ownership, and artifact portability to estimate work needed. - **Permission-First Filtering**: Ensures that only agents with the necessary permissions can be considered. - **Bounded Candidate Search**: Limits the number of candidates to improve efficiency. - **Workload-Sensitive Quotes**: Provides quotes based on workload, ensuring fair resource allocation. - **Auditable Alternatives**: Tracks and records all decisions for transparency. - **Degraded-Route Flags**: Identifies and flags routes that cannot be executed due to degraded conditions. - **State Persistence**: Stores and restores the state of the system to ensure reliability. - **Transport-Neutral A2A Plans**: Provides plans that are independent of the underlying communication protocol. ### Tech Stack: Python ### Why It Matters: Sprix SAGE Router is a critical component of Sprix AI’s research portfolio, focusing on the dynamic decision-making process in A2A networks. By enabling checkpoint-aware reconfiguration, it addresses the challenges of task execution in open networks

### Overview: 快速上手AI理论及应用实战：基础知识、Transformer、NLP、ML、DL、竞赛。含大量注释及数据集，力求每一位能看懂并复现。 ### Key features: - 必备数学基础 - 大模型Large Language Model（Transformer） - NLP（自然语言处理） - ML（机器学习） - DL（深度学习） - DL（深度学习） ### Tech stack: Jupyter Notebook ### Why it matters: 力求每一位能看懂并复现。

Overview: NVIDIA's Transformer Engine is a library designed for accelerating Transformer models on NVIDIA GPUs, offering 8-bit and 4-bit precision training for Hopper, Ada, Blackwell, and Blackwell GPUs, and PyTorch and JAX APIs with mixed-precision support. Key features: - Support for 8-bit and 4-bit precision training - Optimized attention, GEMM, normalization, quantization, and fused Transformer and MoE modules - PyTorch and JAX APIs with mixed-precision support - Support for tensor, sequence, context, and EP, including communication overlap - PyTorch and JAX APIs with autocast-style contexts and configurable low-precision recipes - Support for NCCL-backed EP for PyTorch and JAX, faster MoE routing, and a JAX Flax MoE block Tech stack: Python Why it matters: NVIDIA's Transformer Engine provides a comprehensive solution for accelerating Transformer model training on NVIDIA GPUs, offering optimized performance and reduced memory usage through mixed-precision training and fused kernels, making it an essential tool for developers working with large-scale distributed training. Repository: NVIDIA/TransformerEngine GitHub stars: 3586 GitHub description: A library for accelerating Transformer models on NVIDIA GPUs, including using 8-bit and 4-bit floating point (FP8 and FP4) precision on Hopper, Ada and Blackwell GPUs, to provide better performance with lower memory utilization in both training and inference.

### Overview: This is the Machine Learning course repository for the girafe-ai Machine Learning course, focusing on the first semester. It includes recordings, materials, and assignments from various lectures and seminars. ### Key Features: - **Lectures and Seminars:** Recorded lectures and seminar videos covering topics such as Naive Bayes, kNN, Linear Regression, Linear Classification, SVM, PCA, Trees, Ensembles, Gradient Boosting, Intro to Deep Learning, Backpropagation, and Dropout/Batchnorm. - **Assignments and Homeworks:** Assignments and homeworks for each week, including kNN, Linear Regression, SVM, Tree from scratch, and more. - **GitHub Repository:** The repository is hosted on GitHub, with a primary language of Jupyter Notebook, and has received 3,584 GitHub stars. ### Tech Stack: - **GitHub:** The repository is hosted on GitHub, which is used for version control and collaboration. ### Why It Matters: This course provides a comprehensive overview of Machine Learning concepts and techniques, suitable for developers and learners interested in the field. The recorded lectures and materials ensure that learners can follow along with the course material, making it a valuable resource for anyone looking to enhance their Machine Learning skills.

### Overview: OneCLI is an open-source platform designed to run AI agents as a team, providing secure, sandboxed personal agents for every employee. The platform allows teams to manage and enforce policies across multiple agents, ensuring that each agent operates within a secure environment. ### Key Features: - **Secure Agent Sandboxing**: Agents are isolated and protected by a gateway, ensuring that credentials never leave the sandbox. - **Multiplayer Management**: Teams can manage policies and agents from a central dashboard, making it easy to replicate and scale. - **Multiplayer Collaboration**: Agents can be shared and connected across the workspace, allowing for seamless collaboration and communication. - **Determined Human-in-the-Loop Approvals**: Real-time approvals are available in Slack, ensuring that only authorized actions are taken. - **Policy Enforcement**: A single policy can be enforced across all agents, ensuring consistency and security. ### Tech Stack: - TypeScript ### Why It Matters: OneCLI is essential for teams that need to run autonomous agents for their employees, providing a secure and efficient way to manage and enforce policies across multiple agents. The platform's ability to integrate with identity providers and enforce policies in real-time makes it a valuable tool for modern teams looking to automate and streamline their workflows.

### Overview The **Synthetic Data Vault** (SDV) is a Python library designed to generate synthetic tabular data using machine learning techniques. The SDV offers a variety of models, including classical statistical methods (GaussianCopula) and deep learning methods (CTGAN), allowing you to create synthetic data for single tables, multiple connected tables, or sequential tables. It also provides tools for evaluating and visualizing the quality of the synthetic data, as well as preprocessing, anonymization, and constraint definition. ### Key Features - **Create Synthetic Data**: Generate synthetic data using machine learning algorithms. - **Evaluate and Visualize Data**: Compare synthetic data to real data against various measures and generate quality reports. - **Preprocess, Anonymize, and Define Constraints**: Control data processing, anonymize data, and define business rules. ### Tech Stack Python ### Why It Matters The SDV is essential for anyone looking to create synthetic data for their applications, ensuring that the generated data is realistic and can be used for testing, training models, or simulating scenarios without real data. It simplifies the process of creating and evaluating synthetic data, making it a valuable tool for data scientists and researchers. Repository: sdv-dev/SDV Primary language: Python GitHub stars: 3579

### Overview: makesense.ai is a free online tool designed for labeling photos, making the process of preparing a dataset for small computer vision deep learning projects much easier and faster. It is cross-platform and does not require any installation. ### Key Features: - **Photo Labeling**: Automatically label photos with bounding boxes and suggest labels. - **AI Integration**: Supports YOLOv5, SSD, and PoseNet for enhanced labeling capabilities. - **TypeScript and React/Redux**: Written in TypeScript and based on the React/Redux duo. ### Tech Stack: - **TypeScript**: Primary programming language. - **React/Redux**: Front-end framework. ### Why It Matters: - **Ease of Use**: No installation required, works on any browser. - **Cross-Platform Compatibility**: Works seamlessly across different operating systems. - **Efficiency**: Reduces time spent on manual labeling, making the process of preparing a dataset much faster and easier. ### Repository: Repository: SkalskiP/make-sense GitHub Stars: 3579 GitHub Description: Free to use online tool for labeling photos. [Link](https://makesense.ai) Tech Stack: TypeScript, React/Redux Why It Matters: Free to use online tool for labeling photos, works on any browser, and supports AI integration.

### Overview: This repository contains classic papers and resources on recommendation systems, covering various techniques and models used in the field. It serves as a comprehensive collection of papers, learning materials, and industry insights, aimed at both developers and researchers interested in recommendation systems. ### Key Features: - **Distillation-Based Multi-Task Learning**: Improves reading duration through candidate generation models. - **Personalized Re-ranking for Recommendation**: Enhances recommendation diversity. - **Towards the Next Generation of Pre-ranking System**: Aims to enhance pre-ranking systems. - **Fast Greedy Map Inference for Determinantal Point Processes**: Improves recommendation diversity. - **Learning Tree-based Deep Model for Recommender Systems**: Models large-scale recommendation systems. - **From RankNet to LambdaRank to LambdaMART**: An overview of ranking models. - **Managing Diversity in Airbnb Search**: Techniques for improving search diversity. - **Deep & Cross Network for Ad Click Predictions**: A survey of deep learning models for click predictions. - **Deep Crossing - Web-Scale Modeling without Manually Crafted Combinatorial Features**: Models without manually crafted features. - **Deep Interest Network for Click-Through Rate Prediction**: A survey of deep interest networks. - **Deep Learning based Recommender System**: A survey and new perspectives on deep learning in recommendation systems. ### Tech Stack: Python ### Why it Matters: This repository provides a wealth of information on various recommendation system techniques and models, making it a valuable resource for developers and researchers in the field. It covers a wide range of topics, from basic models to advanced techniques

### Overview Open Intelligent UI is an open-source generative UI framework designed to answer questions and provide explanations, comparisons, and working tools. Built with CopilotKit and AG-UI, it offers a direct text answer, native components, and interactive UIs based on the task at hand. ### Key Features - **Understand**: Explore mechanisms, step through explanations, or change variables. - **Compare**: Examine criteria, assumptions, and tradeoffs side by side. - **Make a tool**: Use a calculator, planner, or interactive model with working controls. ### Tech Stack - TypeScript ### Why It Matters Open Intelligent UI simplifies complex interactions by providing a conversational interface that adapts to different tasks, making it a valuable tool for anyone looking to enhance their productivity or decision-making processes.

Overview: This project is a comprehensive guide to understanding linear algebra through Python and visualization techniques. It covers 66 topics and is designed to make the subject accessible and engaging. Key features: - Interactive Python code examples - Visualizations to explain complex concepts - Step-by-step tutorials Tech stack: Jupyter Notebook Why it matters: This project aims to demystify linear algebra by making it more relatable and understandable through practical applications and visual aids. It is particularly useful for developers looking to enhance their data science skills.

### Overview: EvalScope is a comprehensive and customizable framework designed for efficient evaluation and benchmarking of large language models (LLM), vision language models (VLM), and more. It supports multiple evaluation benchmarks and integrates various evaluation backends for seamless performance testing. ### Key Features: - **📚 Comprehensive Evaluation Benchmarks**: Built-in benchmarks such as MMLU, C-Eval, GSM8K, and more. - **🧩 Multi-modal and Multi-domain Support**: Evaluates models across various domains and types including LLM, VLM, embedding, reranker, and more. - **🧭 System One Evaluation**: Evaluates text-choice decision models with structured probabilities and confidence. - **🚀 Multi-backend Integration**: Seamlessly integrates OpenCompass, VLMEvalKit, and RAGEval for diverse evaluation needs. - **🤖 Agent Evaluation Mode**: Drives benchmarks inside a controlled multi-turn AgentLoop with pluggable strategies and Docker sandbox. - **⚡ Inference Performance Testing**: Provides powerful model service stress testing tools for metrics like TTFT and TPOT. - **📊 Interactive Reports**: Offers a Web Dashboard for model comparison, overview, and detailed analysis. - **⚔️ Arena Mode**: Supports multi-model battles for intuitive ranking and evaluation. - **🔧 Highly Extensible**: Allows developers to add custom datasets, models, and evaluation metrics. ### Why It Matters: EvalScope simplifies the evaluation and benchmarking process, making it easier and more efficient to compare and analyze the performance of large language and vision models across various domains and types.