Projects

Overview: This repository contains coursework, assignments, and slides for the Natural Language Processing Specialization by deeplearning.ai on Coursera. It covers state-of-the-art deep learning techniques for building NLP systems, including sentiment analysis, translation, summarization, and chatbot creation. Key features: - Coursework and assignments - Slides for lectures - State-of-the-art deep learning techniques - Cutting-edge NLP models - Question-answering and sentiment analysis - Language translation - Text summarization - Chatbot creation Tech stack: Jupyter Notebook Why it matters: This specialization equips learners with the skills needed to build advanced NLP systems, making them valuable in fields such as AI, machine learning, and software engineering.

### Overview: This repository contains the implementation of the Swin Transformer, a novel vision transformer architecture. The Swin Transformer is designed to address the challenges of adapting Transformers from language to vision tasks, such as the large variations in the scale of visual entities and the high resolution of pixels in images compared to text. It introduces a hierarchical Transformer with shifted windows, which improves efficiency and flexibility, making it suitable for a wide range of vision tasks, including image classification and dense prediction tasks. ### Key Features: - **Hierarchical Transformer with Shifted Windows**: The Swin Transformer uses a hierarchical architecture with shifted windows to compute representations, which is more efficient and allows for cross-window connections. - **Simplified Self-Attention**: The implementation includes a linear self-attention mechanism, making it compatible with a broad range of vision tasks. - **Performance Superiority**: It surpasses the state-of-the-art by a large margin, demonstrating the potential of Transformer-based models as vision backbones. ### Tech Stack: - **Programming Language**: Python - **Dependencies**: Requires PyTorch and other necessary libraries for the implementation ### Why It Matters: The Swin Transformer's ability to handle large variations in visual entities and high-resolution images makes it a powerful tool for computer vision tasks. Its linear computational complexity and flexibility to model at various scales make it highly adaptable and competitive with other state-of-the-art models in vision tasks. ### GitHub Stars: 863 ### Repository: The repository is maintained by the authors of the Swin Transformer paper, including Ze Liu, Yutong Lin,

### Overview: FastestDet is a lightweight anchor-free target detection algorithm designed to be 250K parameters lighter than Yolo-Fastest, reducing time consumption by 10% compared to Yolo-Fastest. It offers better performance and simpler post-processing, with an AP increase of 0.7 compared to Yolo-Fastest. ### Key features: - **Anchor-Free Architecture**: Eliminates the need for anchor boxes, making it highly efficient and accurate. - **Single Scale Detector Head**: Uses a single scale detector head for improved performance. - **Cross Grid Multiple Candidate Targets**: Utilizes dynamic positive and negative sample allocation for more precise detection. - **Simpler Post-Processing**: Simplifies post-processing, reducing complexity and improving efficiency. ### Tech stack: Python ### Why it matters: FastestDet stands out by offering significant improvements in performance and efficiency, making it a valuable tool for developers looking to enhance their detection algorithms without compromising on accuracy or performance.

### Overview: This package provides utilities for visualizing spaCy models and building interactive spaCy-powered apps with Streamlit. ### Key features: - Visualizers for syntactic dependencies, named entities, text classification, and semantic similarity via word vectors, token attributes. - Building blocks for Streamlit apps that call into Streamlit and set up all required elements. - Includes the English model for quick start. ### Tech stack: Python ### Why it matters: This package simplifies the integration of spaCy models into Streamlit apps, making it easier to create interactive visualizations and analyses.

Overview: Savant is a high-performance, open-source framework for building real-time and video analytics applications on NVIDIA hardware, designed for data centers and edge environments. Key features: - Real-time and high-load computer vision and video analytics applications - Easy to use, flexible, and scalable - Built on DeepStream, providing a high-level abstraction layer for inference pipelines - Compatible with various NVIDIA hardware versions Tech stack: Python, NVIDIA DeepStream, JetPack Why it matters: Savant simplifies the development of complex multimedia AI applications, making it easier and more efficient to build real-time and high-load computer vision and video analytics systems on NVIDIA hardware.

### Overview: MeshGPT is an implementation of MeshGPT, a state-of-the-art method for mesh generation using attention in Pytorch. This project aims to provide a practical and efficient solution for generating 3D models from text descriptions. ### Key features: - **MeshGPT Implementation**: The project provides a complete implementation of the MeshGPT model, showcasing its capabilities in generating 3D meshes from text inputs. - **Pytorch Framework**: Utilizes Pytorch for efficient computation and model training. - **Text Conditioning**: The project includes text conditioning, allowing the generation of 3D assets based on textual descriptions. ### Tech stack: Pytorch ### Why it matters: This project is significant because it demonstrates the practical application of advanced AI techniques like MeshGPT in generating 3D models from text, which can be useful in various applications such as virtual reality, game development, and more. ### Repository: [lucidrains/meshgpt-pytorch](https://github.com/lucidrains/meshgpt-pytorch) ### GitHub Stars: 863 ### GitHub Description: Implementation of MeshGPT, SOTA Mesh generation using Attention, in Pytorch ### Update: - **Marcus** has trained and uploaded a working model to 🤗 Huggingface! ### Appreciation: - **StabilityAI**, **A16Z Open Source AI Grant Program**, and **🤗 Huggingface** for generous sponsorships. - **Marcus** for initial code review and first successful end-to

### Overview: EMOCA is an emotion-driven monocular face reconstruction system that accurately captures and reconstructs facial expressions from single images, enhancing the emotional content of the input image. This project is part of the CVPR 2022 paper "EMOCA: Emotion-Driven Monocular Face Capture and Animation." ### Key Features: - **Monocular Face Reconstruction:** EMOCA reconstructs a 3D face from a single image, capturing facial expressions accurately. - **Emotion-Driven Reconstruction:** The system focuses on capturing the emotional state of the input image, enhancing realism and authenticity. - **State-of-the-Art Performance:** EMOCA sets new standards in reconstructing highly emotional images in-the-wild, surpassing previous methods. ### Tech Stack: - **Programming Language:** Python - **Dependencies:** PyTorch, PyTorch Lightning ### Why It Matters: EMOCA is a significant advancement in the field of monocular face reconstruction, particularly in capturing and reconstructing facial expressions accurately. This system not only improves the realism of the reconstructed face but also enhances the emotional content, making it a valuable tool for applications such as facial animation and emotion analysis. ### Repository: The official repository for EMOCA is located at <https://github.com/radekd91/emoca>. It is hosted on GitHub and has received 862 GitHub stars, indicating its popularity and utility in the community. ### Installation and Usage: To use EMOCA, you can install the official implementation using the following commands: ```bash git clone https

Overview: ChatPDF is a lightweight implementation of GraphRAG designed for local deployment with native Python, supporting various LLM models and file formats, and optimizing the RAG accuracy. Key features: - Supports local mode for relationship graph-based document question answering - Integrates OpenAI API, Deepseek API, and Ollama API for seamless integration - Supports multiple open-source LLM models, including Qwen/DeepSeek - Supports various file formats, including PDF, docx, markdown, and txt - Optimizes RAG accuracy with features like Chinese chunk cut optimization, embedding optimization, and reranker model tuning - Uses Gradio for real-time chat interface Tech stack: Python Why it matters: ChatPDF is a practical solution for developers looking to integrate local LLMs with RAG, providing a user-friendly interface and optimized performance. Repository: shibing624/ChatPDF Primary language: Python GitHub stars: 862 GitHub description: RAG for Local LLM, chat with PDF/doc/txt files, ChatPDF.

Overview: AgentForge is an extensible AGI framework designed for rapid development, testing, and iteration of AI-powered autonomous agents and cognitive architectures. It supports flexible Agents, declarative Cogs, and optional Memory, making it compatible with various LLM models and easy to use. Key features: - Declarative Cogs for orchestration of multi-agent workflows. - Customizable Agents using YAML prompt templates. - Integrated Memory for context-aware interactions. - Personas for configuring agent identity. - Dynamic Prompt Templates for flexible prompts. - LLM Agnostic support for different LLMs. - On-The-Fly Prompt Editing for real-time modifications. - Open-Source Model Support for local models via Ollama and LMStudio. Tech stack: Python. Why it matters: AgentForge simplifies the development of AI-driven autonomous agents and cognitive architectures, making it accessible for developers of all skill levels. Repository: DataBassGit/AgentForge GitHub stars: 860 GitHub description: Extensible AGI Framework

Overview: ControlVideo is an official PyTorch implementation of the "ControlVideo" paper, which aims to directly inherit the high-quality and consistent generation capabilities of ControlNet without any fine-tuning. This project is designed to be a versatile tool for generating controllable videos from text descriptions. Key features: - Directly inherits the high-quality and consistent generation capabilities of ControlNet - Supports various structure sequences for better control - Compatible with different versions of ControlNet, including 1.0 and 1.1 - Supports efficient long-video synthesis for longer video generation Tech stack: Python Why it matters: ControlVideo simplifies the process of generating controllable videos by leveraging the strengths of existing models like Stable Diffusion and ControlNet, making it an accessible solution for developers looking to enhance their text-to-video generation capabilities without the need for extensive fine-tuning.

### Overview: Catalyst is a high-performance C# Natural Language Processing library designed for speed. Inspired by spaCy's design, it provides pre-trained models, out-of-the-box support for training word and document embeddings, and flexible entity recognition models. ### Key Features: - **Fast, Modern Pure-C# NLP Library**: Targeting .NET 9 and .NET 10, running on Windows, Linux, macOS, and ARM. - **Cross-Platform Support**: Non-destructive tokenization, over 99.9% RegEx-free, over 1M tokens/s on a modern CPU. - **Named Entity Recognition**: Supports gazetteer, rule-based, and perceptron-based models. - **Pre-Training Models**: Based on the Universal Dependencies project. - **Custom Models**: Supports learning Abbreviations and Senses. - **Embeddings**: Out-of-the-box support for FastText and StarSpace embeddings (coming soon). - **Language Detection**: FastText or cld3 language detection. - **Efficient Serialization**: Binary serialization using MessagePack. - **Pre-built Models**: Language packages available for NuGet. ### Tech Stack: Catalyst is built using C# and leverages modern C# features for performance and ease of use. ### Why It Matters: Catalyst is essential for developers looking to integrate advanced NLP capabilities into their .NET applications efficiently. Its speed and flexibility make it a valuable tool for a wide range of applications, from language translation and text analysis to sentiment analysis and more. ### GitHub Stars: 86

### Overview: Elephant-Goldfish Model (EGM) is an open-source plugin designed for Google to facilitate Spec-Driven Development with AI coding agents. It leverages the Elephant metaphor to draft specifications, while the Goldfish metaphor ensures designs are rigorously tested through isolated subagents. ### Key Features: - **Brainstorm**: Zero-doc ideation tool using a multi-lens Goldfish shootdown swarm to stress-test early concepts. - **Draft**: Iterative, section-by-section markdown design specification compiler. - **Review**: Multi-stage verification pipeline orchestrator. - **Review Loop**: Concurrent review subagent dispatch engine. - **Review Resolve**: Interactive triage and automated patch application engine. - **Drift Report**: Sequential AST scanner detecting divergence between checked-in specs and live code. - **Bootstrap**: Recursive bottom-up codebase documentation generator and sentinel-block upserting engine. ### Tech Stack: Python ### Why It Matters: EGM is a powerful tool for developers aiming to create high-quality, maintainable code through a structured and rigorous development process. By leveraging the Elephant and Goldfish metaphors, it ensures that AI agents are used responsibly and effectively, leading to more efficient and reliable software development practices. ### Repository: **google/elephant-goldfish-model** **GitHub Stars: 859** **GitHub Description: An open-source method and skill set created at Google for spec-driven development with LLMs**

Overview: agentcookie is a solution designed to keep your agent sessions synchronized across different devices, ensuring seamless and continuous access without requiring manual authentication. Key features: - Synchronization of agent sessions across different devices - Continuous synchronization with no need for manual authentication - Encryption over Tailscale for secure data transfer - Integration with various Chromium-based automation tools Tech stack: Go Why it matters: This tool addresses the common issue of needing to log in twice when using different devices, providing a more seamless and efficient experience for remote work and collaboration. Repository: mvanhorn/agentcookie Primary language: Go GitHub stars: 859 GitHub description: Your agent runs on a Mac that isn't your daily driver. agentcookie keeps its sessions in sync with the Mac you actually browse on, continuously, encrypted over Tailscale, so OpenClaw, Hermes, or any other agent runtime wakes up authenticated. macOS, peer-to-peer, no cloud middleman.

### Overview: Agentfiles is an AI skills manager for Obsidian, allowing users to browse, create, and manage skills across multiple coding agents such as Claude Code, Cursor, Codex, Windsurf, and 17 other tools. It provides a unified interface for skills, commands, and agents, enhancing productivity and organization. ### Key features: - **Browse Skills, Commands, and Agents**: Easily find and manage skills, commands, and agents across various coding tools. - **Search by Name or Content**: Perform deep searches for skills and commands. - **Create New Skills**: Use a wizard to create skills with a pick tool, type, and name. - **Inline Editing**: Edit skills directly within the vault with markdown preview and save functionality. - **Marketplace**: Install skills from the skills.sh marketplace. - **Conversations**: Browse Claude Code session history, search, tag, and export to vault. - **Dashboard**: View usage analytics, burn rate, context tax, and health metrics. ### Tech stack: TypeScript ### Why it matters: Agentfiles simplifies the management of skills across multiple coding tools, enhancing productivity and organization within Obsidian.

Overview: This PHP library is designed to detect the language of a given text string, supporting 110 languages. Key features: - Detects the language of a text string - Parses training text into N-grams for database generation - Uses a database file built from training text for language detection - Supports multiple languages Tech stack: PHP Why it matters: This library simplifies language detection in PHP applications, making it easier to integrate language detection functionality into projects that require it. Repository: patrickschur/language-detection GitHub stars: 858

Overview: LLVIP is a visible-infrared paired dataset designed for low-light vision research. It contains 30,976 images, enabling researchers to study pedestrian detection and other low-light vision tasks. Key features: - Visible-infrared paired dataset for low-light vision - Pedestrian detection models released - Kaggle competition based on part of the dataset - Updated annotations and raw data for further research - Pre-trained pix2pixGAN released Tech stack: Jupyter Notebook Why it matters: LLVIP provides a comprehensive resource for researchers working on low-light vision, offering a valuable dataset and tools for pedestrian detection and other related tasks.

Overview: Council is an open-source platform designed for the rapid development and robust deployment of customized generative AI applications using Python. Key features: - Unified LLM Interface - Provider Flexibility - Usage Monitoring - Configuration Management - Error Handling Tech stack: Python Why it matters: Council simplifies the integration and monitoring of large language models, making it an essential tool for enterprise-grade AI applications. Repository: chain-ml/council GitHub stars: 857

Overview: ESP-AI 是一个简单且成本最低的 AI 接入解决方案，适用于任何硬件设备。它通过将 AI 直接入口集成到项目中，无需影响现有项目，使开发板上的对话功能成为可能。 Key features: - 自定义离线唤醒词，支持多种唤醒方式（语音、按钮、串口、Tianwen ASRPro） - 完整的对话链路：IAT（ASR）→ LLM/RAG → TTS - 快速响应算法，设计用于平衡服务成本并提供最快的响应时间 - 支持对话中断 - 可配置 - 插件基础，支持与任何 LLM/TTS/IAT 进行插件集成 - 一对一服务与客户端关系，每个客户端独立配置 - 支持认证连接 - 全链路流式数据交互 - 开发者平台提供：免费服务、可视化配置等 - 提供客户端配置网页 - 高并发场景支持（需要 Nginx 进行负载均衡） - 出厂即用 Tech stack: Node.js 和 Arduino/IDF Why it matters: ESP-AI 提供了一个简单且成本最低的解决方案，使开发者能够轻松地将 AI 直接入口集成到他们的硬件设备中，从而实现对话功能。这不仅简化了开发过程，还提高了用户体验。 Repository: wangzongming/esp-ai GitHub stars: 857 GitHub description: The simplest and lowest cost

Overview: This project presents a PyTorch implementation of the AdaptSegNet method for adapting semantic segmentation from synthetic to real datasets, specifically focusing on the 2018 CVPR paper "Learning to Adapt Structured Output Space for Semantic Segmentation". Key features: - PyTorch implementation - Adaptation from synthetic to real datasets - Results ranked 3rd in the VisDA Challenge 2018 Tech stack: Python Why it matters: This method demonstrates the potential of adapting semantic segmentation models to real-world datasets, improving their performance and applicability. Repository: wasidennis/AdaptSegNet Primary language: Python GitHub stars: 857 GitHub description: Pytorch implementation of our method for adapting semantic segmentation from the synthetic dataset (source domain) to the real dataset (target domain). Based on this implementation, our result is ranked 3rd in the [VisDA Challenge](http://ai.bu.edu/visda-2017/). Contact: Yi-Hsuan Tsai (wasidennis at gmail dot com) and Wei-Chih Hung (whung8 at ucmerced dot edu) ## Paper [Learning to Adapt Structured Output Space for Semantic Segmentation](https://arxiv.org/abs/1802.10349) <br /> [Yi-Hsuan Tsai](https://sites.google.com/site/yihsuantsai/home)\*, [Wei-Chih Hung](https://hfslyc.github.io/)\*, [Samuel Schul

### Overview: Surogate is a comprehensive toolkit designed to train and serve large language models (LLMs) at extreme speeds and massive throughput. It leverages NVIDIA hardware to optimize both training and serving processes, offering a single, unified solution for developers. ### Key Features: - **Native C++/CUDA LLM Training and Serving**: Supports GGUF, NVFP4, BF16, and FP8 formats. - **Training Speed**: Native GGUF decoding at 802 tokens/s on one RTX 5090, serving 16 users at 7× llama.cpp's aggregate decode throughput on eight cards. - **Serving Speed**: Serving at 7.0× throughput compared to llama.cpp on 16 users. - **Model Optimization**: Supports QLoRA, GRPO, and DPO optimization techniques. - **Training Throughput**: 53,900 training tokens/s on a single H100, 136,200 on four RTX 5090s. ### Tech Stack: - **Engine**: C++ / CUDA ### Why It Matters: Surogate simplifies the training and serving process, making it easier for developers to scale their models without compromising performance. Its focus on speed and efficiency ensures that models can be fine-tuned and deployed quickly, making it an essential tool for developers working on large language models.