Projects

### Overview: TrafficLab is a digital-twin style traffic visualization project that allows anyone to create advanced computer vision demonstrations using only mp4 CCTV footage and its Google Maps location. The tool is accessible to students, investigators, and enthusiasts without access to camera calibration and high-quality satellite imagery. ### Key Features: - **Calibration:** Establishing a two-way projection between any CCTV and its satellite map, with support for custom SVG. - **Inference:** Easily swap object detection models and object tracker, and control various arguments. - **Visualization:** A "digital twin" experience with synchronized view of CCTV footage and 3D bounding boxes. - **Accessibility:** Anyone can create a fancy digital twin demo with just access to mp4 CCTV footage and its location on Google Maps. ### Tech Stack: - Python ### Why It Matters: TrafficLab provides a powerful and accessible platform for students, investigators, and enthusiasts to demonstrate advanced computer vision techniques, especially in scenarios where camera calibration and high-quality satellite imagery are not available. This tool democratizes the use of computer vision in traffic analysis, making it more accessible and impactful. ### Repository: Repository: duy-phamduc68/TrafficLab-3D Primary Language: Python GitHub Stars: 491 GitHub Description: Create a digital-twin style traffic visualization using only mp4 CCTV footage and its Google Maps location. ### Complementary Resources: - [TrafficLab 3D v1.1](https://github.com/duy-phamduc68/TrafficLab-3D/releases/tag/v1.

Overview: Eidolon is a pluggable Agent SDK and enterprise-ready deployment server designed for building and deploying agent-based services. Key features: - Easy to deploy - Simple agent-to-agent communication - Painless component customization and upgrade Tech stack: Python Why it matters: Eidolon simplifies the process of building and deploying agent-based services, making it easier for developers to adapt to the rapidly changing AI landscape without vendor lock-in.

### Overview: This repository is a collection of cutting-edge projects and papers focused on Speech Synthesis, Text-to-Speech (TTS), Singing Voice Synthesis (SVS), Voice Conversion (VC), and Singing Voice Conversion (SVC) related to the field of audio synthesis and voice conversion. It includes a diverse range of research and projects, covering various techniques and applications in speech synthesis, music synthesis, and voice conversion. The repository is updated regularly with the latest advancements and includes detailed demos and code snippets for each project. ### Key features: - **Speech Synthesis (TTS)** - **Singing Voice Synthesis (SVS)** - **Voice Conversion (VC)** - **Singing Voice Conversion (SVC)** ### Tech stack: Python, PyTorch, TensorFlow, and other relevant libraries and frameworks. ### Why it matters: This repository is essential for developers and researchers interested in advancing the fields of speech synthesis, voice conversion, and related applications. It provides a comprehensive overview of the latest research in the field, including detailed projects, papers, and demos, making it a valuable resource for anyone looking to stay updated with the latest developments in audio synthesis and voice conversion.

### Overview: Tool learning with large language models (LLMs) has emerged as a promising approach for enhancing LLM capabilities to tackle complex problems. This survey aims to provide a comprehensive overview of existing works on tool learning with LLMs, focusing on the benefits of tool integration and the implementation of tool learning paradigms. ### Key Features: - **Tool Learning with LLMs Survey** - **Survey on Tool Learning with LLMs** - **Tool Learning with LLMs: A Comprehensive Survey** ### Tech Stack: - **Python** - **Git** - **GitHub** - **Frontiers of Computer Science (FCS)** ### Why It Matters: This survey aims to bridge the fragmented literature on tool learning with LLMs, providing a systematic understanding of the benefits and implementation methods. It is essential for researchers and developers looking to leverage LLMs for complex problem-solving tasks.

Overview: The `dots.llm1` model is a large-scale Multi-Head Attention (MoE) model with 14B activated parameters out of 142B total parameters. It is pre-trained on high-quality corpus and fine-tuned for instruction-tuning, achieving performance comparable to Qwen2.5-72B. The model is open-sourced, providing intermediate checkpoints for further research into large language model learning dynamics. Key features: - Type: MoE model with 14B activated and 142B total parameters - Training Stages: Pretraining and SFT - Architecture: Multi-head Attention with QK-Norm, fine-grained MoE, and 2 shared experts - Number of Layers: 62 - Number of Attention Heads: 32 - Supported Languages: English, Chinese - Context Length: 32,768 tokens - License: MIT Tech stack: The model is open-sourced and available on Hugging Face. Why it matters: The model is an open-source implementation of a large-scale MoE model, providing insights into the learning dynamics of large language models through intermediate checkpoints spanning the entire training process. This facilitates future research and can be used for further development and research into large language model performance and efficiency.

### Overview: Command Line Artificial Intelligence `CLAI` is an open-sourced project from IBM Research aimed to bring the power of AI to the command line interface. Using CLAI, users can access a wide range of skills that enhance their command line experience. ### Key Features: - **CLI Skills Integration**: CLAI integrates AI capabilities directly into the command line, making complex tasks easier and more intuitive. - **Command Line Enhancements**: Users can perform advanced AI tasks such as sentiment analysis, code completion, and more directly from the terminal. - **Scalability and Flexibility**: CLAI is designed to be easily scalable and can be integrated into various environments, including z/OS. ### Tech Stack: Python ### Why it Matters: CLI skills integration opens up new possibilities for developers and users alike, making everyday tasks more efficient and enjoyable. CLAI's ability to enhance the command line interface is particularly valuable in environments where productivity is critical, such as in the enterprise or for developers working on complex systems. ### Repository: IBM/clai ### Primary Language: Python ### GitHub Stars: 493 ### GitHub Description: Command Line Artificial Intelligence `CLAI` is an open-sourced project aimed to bring the power of AI to the command line. Using CLAI, users of Bash can access a wide range of skills that will enhance their command line experience. This repository contains the source code and documentation to get you started. ### Getting Started: - **Home**: See examples of CLAI in action without installing locally. - **More**: A quick overview about the

### Overview: This project provides Python scripts for performing object detection using the YOLOv8 model in ONNX format. The scripts are designed to process images and videos efficiently, leveraging the power of the YOLOv8 model for accurate object detection. ### Key Features: - **Image Detection**: The scripts can detect objects in images. - **Webcam Detection**: The scripts can detect objects in real-time using a webcam. - **Video Detection**: The scripts can detect objects in videos. ### Tech Stack: - Python - ONNX (Open Neural Network Exchange) ### Why It Matters: The project is crucial for developers and researchers looking to integrate object detection capabilities into their applications using the YOLOv8 model. The scripts are designed to be efficient and accurate, making them suitable for various applications requiring object detection. ### GitHub Stars: 492 ### Installation: ```shell git clone https://github.com/ibaiGorordo/ONNX-YOLOv8-Object-Detection.git cd ONNX-YOLOv8-Object-Detection pip install -r requirements.txt ``` ### Requirements: - Ensure you have Python installed. - Install NVIDIA GPU support if available, or use the CPU version. - Convert the YOLOv8 model to ONNX format using the provided script or Google Colab notebook. ### ONNX Model: The project includes the original YOLOv8 model and the converted ONNX model. The original YOLOv8 model can be found in the YOLOv8 repository, and

Overview: This repository is an official PyTorch implementation of the "Camera Distance-aware Top-down Approach for 3D Multi-person Pose Estimation from a Single RGB Image" paper presented at ICCV 2019. It includes RootNet and is compatible with various publicly available datasets such as Human3.6M, MPII, MS COCO 2017, MuCo-3DHP, MuPoTS-3D, and 3DPW. Key features: - PyTorch implementation of the paper's approach - Flexible and simple codebase - Compatibility with multiple datasets including Human3.6M, MPII, MS COCO 2017, MuCo-3DHP, MuPoTS-3D, and 3DPW - Includes visualization code for human pose estimation Tech stack: Python Why it matters: This repository provides a practical and efficient solution for 3D multi-person pose estimation using a single RGB image, making it valuable for researchers and developers working on similar projects.

Overview: HistomicsTK is a Python toolkit designed for the analysis of digital pathology images, offering a range of algorithms for tasks such as color normalization, deconvolution, nuclei segmentation, and feature extraction. Key features: - Color normalization - Color deconvolution - Nuclei segmentation - Feature extraction Tech stack: Python Why it matters: HistomicsTK provides a comprehensive solution for the analysis of digital pathology images, addressing the needs of both pathologists and algorithm researchers. It bridges the gap in open-source tools for managing, visualizing, and analyzing digital pathology data, making it an essential tool for the field of computational pathology.

### Overview: SC-Depth is a PyTorch Lightning implementation of three versions of **self-supervised learning for monocular depth estimation**. The repository includes the SC-DepthV1, SC-DepthV2, and SC-DepthV3 models, each designed to address specific challenges in unsupervised monocular depth estimation. The models are trained using geometry consistency loss and self-discovered masks to improve accuracy and consistency. ### Key Features: - **SC-DepthV1**: Geometry consistency loss for scale-consistent depth prediction over time. - **SC-DepthV2**: Auto-recitify network (ARN) for handling large relative rotation motions in hand-held camera videos. - **SC-DepthV3**: Robust learning framework for accurate and sharp monocular depth estimation in dynamic scenes. ### Tech Stack: - **Python**: The primary programming language used for implementation. - **PyTorch Lightning**: The framework used for the PyTorch Lightning implementation. ### Why It Matters: SC-Depth models provide a significant improvement in accuracy and consistency for monocular depth estimation, making them valuable tools for applications such as autonomous driving, augmented reality, and robotics. The models are validated on a variety of datasets, including static and dynamic scenes, demonstrating their effectiveness in different scenarios. ### Repository: - **Repository**: JiawangBian/sc_depth_pl - **GitHub Stars**: 491 - **Primary Language**: Python - **GitHub Description**: PyTorch Lightning implementation of SC-Depth (V1, V2, and V3) for self

Overview: NextGenAI is a student-led organization focused on empowering high schoolers with ethical AI skills and fostering global collaboration through seminars, internships, and innovative projects. Key features: - Accelerated AI Learning bootcamps - Innovation & Internships - Global Outreach & Connect - AI Social Impact projects Tech stack: HTML, CSS, JavaScript Why it matters: NextGenAI aims to prepare the next generation of responsible AI leaders by equipping them with essential skills and fostering a global community of ethical AI practitioners.

### Overview: This repository contains best practices, reference architectures, and examples for distributed AI training and inference on AWS. It covers various model families, training frameworks, and serving engines, providing practical guidance for operating clusters on Amazon SageMaker HyperPod, AWS ParallelCluster, AWS Parallel Computing Service, and Amazon Elastic Kubernetes Service. ### Key features: - **Cluster Reference Architectures**: Includes CloudFormation, Terraform, and other scripts for creating Amazon Machine Images (AMI) and deploying clusters. - **Training Examples**: Covers different model families and sizes, training frameworks (PyTorch DDP/FSDP, Megatron-LM, NeMo), and serving engines (vLLM, SGLang, NVIDIA Dynamo). - **Validation Tools**: Provides environment and cluster health validation tools. - **Observability**: Includes monitoring, metrics exporters, and profiling tools. - **Micro-benchmarks**: Offers micro-benchmarks for performance testing. ### Tech stack: Shell ### Why it matters: This repository is essential for developers looking to deploy and optimize distributed AI models on AWS. It provides practical, scalable solutions for training and inference, ensuring efficient and reliable performance across various frameworks and architectures.

Overview: This project implements the KinectFusion approach in modern C++14 and CUDA, leveraging graphics hardware for real-time dense surface mapping and tracking. Key features: - Real-time fusion of depth scans and corresponding RGB color images - Easy to use, modern C++14 interface - Export of the resulting volume as pointcloud - Export also as dense surface mesh using the MarchingCubes algorithm - Functions for easy export of pointclouds and meshes into the PLY file format - Retrieval of calculated camera poses for further processing Tech stack: Cuda Why it matters: This implementation allows for real-time dense surface mapping and tracking using modern C++14 and CUDA, making it suitable for applications requiring real-time performance and efficient GPU utilization. Repository: chrdiller/KinectFusionLib GitHub stars: 491

Overview: This is a professional website and resume builder application that allows users to showcase their skills and experience. It includes a built-in AI assistant and a responsive design for easy access on both mobile and desktop devices. Key features: - A fully functional Angular 21 (LTS) application with a customizable template and content. - Responsive design for mobile and desktop devices. - Internationalization for English and Portuguese. - Mobile navigation sharing. - A Wiki documentation and community chat for support. Tech stack: TypeScript, Angular 21 (LTS), Firebase, Google Analytics, schema.org, structured data. Why it matters: This application provides a professional and easy-to-use website and resume builder, making it a valuable tool for showcasing skills and experience.

Overview: This project provides a guide on how to implement natural language semantic search for arbitrary objects using deep learning techniques. Key features: - Implementing a semantic code search system - Using deep learning methods for efficient object search - Providing a step-by-step tutorial with Jupyter notebooks Tech stack: Python, Jupyter Notebook Why it matters: This project offers a practical approach to creating a semantic code search system, which is essential in fields like software development and data analysis where precise object identification is crucial.

### Overview: UniSpeech is a family of large-scale pre-trained models designed for speech-related tasks, including ASR, speaker diarization, and speaker identification. These models are trained using self-supervised learning techniques and can be fine-tuned for various applications. ### Key features: - **WavLM**: A large-scale self-supervised pre-training model for full-stack speech processing. - **UniSpeech**: Unified pre-training for self-supervised learning and supervised learning for ASR. - **UniSpeech-SAT**: Universal speech representation learning with speaker-aware pre-training. - **ILS-SSL**: Self-supervised learning for speech recognition with intermediate layer supervision. ### Tech stack: Python ### Why it matters: UniSpeech models are highly effective for speaker-related tasks, particularly due to their ability to learn speaker-specific representations. This makes them invaluable for applications such as speaker diarization, speaker identification, and speech recognition. The models are pre-trained on large datasets and fine-tuned on specific tasks, ensuring high performance and robustness.

### Overview: A curated collection of learning resources for Generative AI, Machine Learning, Agentic AI, and related topics. ### Key features: - Comprehensive guides on supervised, unsupervised, and reinforcement learning. - Explainer articles on deep learning, natural language processing, computer vision, and generative adversarial networks. - Documentation on dimensionality reduction, clustering algorithms, Bayesian inference, and time series analysis. - A curated list of AI/ML resources, including courses, certifications, books, and research papers. ### Tech stack: TypeScript ### Why it matters: Stay updated with the latest in AI and ML, from foundational concepts to advanced techniques and practical applications.

Overview: OmniVCus is a project that focuses on the customization of video content using subject-driven approaches and multimodal control conditions. This work was presented at NeurIPS 2025 and aims to improve video customization by providing a data construction pipeline and a diffusion Transformer model. Key features: - Data construction pipeline for creating subject-driven video customization data pairs. - Multimodal control conditions including subjects, depth, mask, motion, etc. - Re-trained model based on public codes and datasets. Tech stack: Python Why it matters: OmniVCus provides a practical solution for enhancing video customization by incorporating subject-driven approaches and multimodal control conditions, which can be particularly useful in applications like Adobe Firefly Video Foundation Model training and testing. Repository: caiyuanhao1998/Open-OmniVCus Primary language: Python GitHub stars: 490 GitHub description: OmniVCus: Feedforward Subject-driven Video Customization with Multimodal Control Conditions (NeurIPS 2025)

### Overview: This project introduces Compositional Visual Generation with Composable Diffusion Models, focusing on the use of conjunction and negation operators to generate compositional images using conditional diffusion models. The method is applied to improve the creativity and diversity of images generated by models like Point-E and Stable Diffusion. ### Key Features: - Compositional generation using conjunction and negation operators - Conditional diffusion models for text-conditioned generation - Training and inference code for text-conditioned diffusion models ### Tech Stack: - PyTorch - GLIDE - Improved-Diffusion ### Why It Matters: This approach enhances the creativity and diversity of generated images by leveraging compositional operators, making the models more versatile and capable of producing more intricate and varied results.

### Overview: PyKale is a knowledge-aware machine learning library designed for interdisciplinary research. It bridges data, software, and end users, making machine learning more accessible. PyKale focuses on multimodal learning and transfer learning, enabling efficient and sustainable AI systems. The library is part of the PyTorch ecosystem and is known for its green machine learning principles, standardization, and minimalism. ### Key features: - **Pipeline-based API**: Unified interface for multimodal learning and transfer learning. - **Multimodal Support**: Compatible with graphs, images, and videos. - **Green Machine Learning**: Reduces repetitions, reuses existing resources, and recycles learning models across areas. ### Tech stack: - Python ### Why it matters: PyKale is a crucial tool for interdisciplinary research, enabling more accessible and sustainable machine learning practices. Its focus on knowledge-aware machine learning makes it essential for applications like bioinformatics, graph analysis, image/video recognition, and medical imaging. ### Repository: Repository: pykale/pykale GitHub stars: 490 ### Primary language: Python ### GitHub description: "Knowledge-Aware machine LEarning (KALE): accessible machine learning from multiple sources for interdisciplinary research, part of the 🔥PyTorch ecosystem. ⭐ Star to support our work!"