Projects

### Overview: Giselle is an open-source AI workflow tool designed for local and self-hosted use, facilitating seamless human-AI collaboration. ### Key features: - **⚡ GitHub AI Operations** - Automates issues, PRs, and dependencies. - **Visual Agent Builder** - Streamlines the creation of AI agents with intuitive visual tools. - **Multi-Model Composition** - Allows for the integration of multiple AI models for complex tasks. - **Knowledge Store** - A repository for storing and accessing knowledge. - **Team Collaboration** - Enhances collaboration among team members. - **Template Hub** - Provides a library of pre-built AI agent templates. ### Tech stack: TypeScript ### Why it matters: Giselle empowers developers and teams to build, test, and deploy AI workflows locally, making it an essential tool for integrating AI into their projects without the need for cloud services or expensive infrastructure.

### Overview: This repository contains a curated list of Korean speech recognition resources for developers, including the error rate (Character Error Rate) of Speech Recognition APIs on public datasets. ### Key features: - **API List**: A collection of Korean speech recognition APIs including those from platforms like OpenAI, Google, ETRI, Naver, and Return Zero. - **Error Rate Measurement**: The error rate (Character Error Rate) of Speech Recognition APIs measured on public datasets. - **Introduction to Korean Speech Recognition Evaluation**: An overview of Korean speech recognition evaluation methods. ### Tech stack: None specified ### Why it matters: This repository provides a comprehensive list of Korean speech recognition APIs along with their error rates, making it an essential resource for developers looking to evaluate and select the best Korean speech recognition solutions.

Overview: 这个项目是一个基于深度学习算法实现的中文问答系统，旨在通过双向长短期记忆网络模型在多个句子中找到给定问题的答案。尽管存在一些不恰当的实现和模型选择，但使用互联网第三方资源进行验证，MRR可达0.75以上。 Key features: - 使用双向长短期记忆网络模型 - 通过jieba分词和word embedding编码进行文本处理 - 评估脚本提供MRR、MAP和ACC@1的评估结果 Tech stack: Python Why it matters: 尽管存在一些不恰当的实现和模型选择，该项目展示了深度学习在中文问答系统中的应用潜力，为后续研究提供了宝贵的数据和模型参考。

### Overview: This repository contains a collection of Google Colab Notebooks designed for various projects and scripts. The notebooks cover topics such as image generation, video creation, and text-to-video conversion, utilizing advanced models like ComfyUI and LTX. ### Key Features: - **Image Generation**: Includes notebooks for generating images using Qwen-Image and ComfyUI. - **Video Creation**: Features notebooks for video generation using ComfyUI and LTX. - **Text-to-Video Conversion**: Includes notebooks for converting text to video using LTX. ### Tech Stack: The primary language used is Jupyter Notebook, leveraging advanced models for text, image, and video generation. ### Why It Matters: These notebooks provide a comprehensive set of tools for developers and researchers to experiment with and enhance their projects, leveraging cutting-edge AI techniques in a reproducible and accessible manner.

Overview: AvatarAI is an open-source AI avatar platform that allows users to create lifelike conversations with any face, voice cloning, and real-time lip-sync video. It includes features like zero-shot voice cloning, multi-LLM support, and local 100% deployment. Key features: - Zero-shot voice cloning - Any face, any language support - Token-streaming pipeline - Barge-in feature - 100% local mode - Multi-LLM support - AWS GPU deployment Tech stack: Python, Next.js, FastAPI, TypeScript, Docker, PostgreSQL, CUDA, Redis, Docker Compose, AWS, AWS GPU deployment. Why it matters: AvatarAI is the most complete open-source AI avatar system, offering real-time conversations with lip-sync video, zero-shot voice cloning, and local 100% deployment.

### Overview Voice Overlay for iOS is a tool that helps users grant their voice permission and input it as text in a customizable user interface. It integrates seamlessly with the native `SFSpeechRecognizer` to perform speech-to-text conversion, ensuring a polished user experience. ### Key Features - **Voice Permission**: Automatically prompts users for voice permission. - **Text Input**: Converts spoken input into text. - **Customization**: Allows for easy customization of the UI and behavior. - **Permissions Handling**: Manages necessary permissions for speech recognition. ### Tech Stack - **Swift**: The primary programming language for development. - **Carthage**: Used for dependency management, ensuring compatibility with Swift 4 and later. - **CocoaPods**: Compatible with InstantSearchVoiceOverlay, ensuring swift compatibility. - **SFSpeechRecognizer**: Native framework for speech-to-text conversion. ### Why It Matters Voice Overlay simplifies the process of integrating voice recognition into iOS applications, making it easier for developers to implement voice-to-text functionality without the need for extensive knowledge of speech recognition APIs. ### Repository The repository is maintained by the Algolia team and is open-source, with 556 GitHub stars. It is compatible with Swift 4 and later, and supports Carthage and CocoaPods. ### Installation - **Swift Package Manager**: Add the package to your project and use SwiftPM to install. - **CocoaPods**: Add the pod to your Podfile and install. - **Carthage**: Add the Carthage dependency to your Cartfile. ### Usage

Overview: 这个项目是追一科技首届中文NL2SQL挑战赛的决赛方案，通过Jupyter Notebook的形式呈现，展示了模型的构建和训练过程。 Key features: - Model 1：将问题与表头顺序连接，每个列前添加特殊标记。 - Model 2：预测条件值，通过枚举候选组合进行二分类问题的解决。 - 模型训练中的探索尝试：优化策略和成功、不成功的想法。 Tech stack: TensorFlow, Keras Why it matters: 该项目展示了如何将自然语言转换为SQL语句，为解决中文领域的NLP任务提供了有价值的参考和启发。

Overview: This repository demonstrates the implementation of a vehicle detection and tracking system using a camera mounted inside a self-driving car. The system utilizes a pre-trained TensorFlow Object Detection API model to detect vehicles and a Kalman-filtering-based tracker to track their positions. The code is designed for simplicity and readability, making it accessible to developers, researchers, and engineers. Key features: - Vehicle detection using a pre-trained SSD MobileNet model - Kalman-filtering-based vehicle tracking - Simple and readable code implementation Tech stack: Python Why it matters: This project provides a practical framework for developers to quickly test and compare different vehicle detection and tracking algorithms, focusing on ease of use and readability.

### Overview: This repository is a comprehensive guide for developers looking to build, deploy, and understand Large Language Model (LLM) agents. It includes 60+ agent skeletons, tutorials, ecosystem guides, and evaluation tools. ### Key Features: - **Agent OS**: Modern agent operating system, identity, memory, skills, MCP, safety, observability. - **Provider Ecosystem**: Adapters for 24+ LLM providers (frontier APIs, fast inference, marketplaces, enterprise clouds, specialty, local runtimes). - **Skills Ecosystem**: Design guide, taxonomy, maturity model, security checklist, and a curated skill catalog. - **Prompt Engineering**: Agent prompt patterns, instruction hierarchy, context engineering, prompt-injection defense. - **Coding-Agent Workflows**: Templates, blueprints, runnable adapters, and curated examples for Claude Code, Cursor, Codex, Aider, Cline, and custom runtimes. - **Design Docs**: Agent / technical design docs, ADR guide, design reviews, rollout plans, the `DESIGN.md` machine-readable spec. - **Curated Catalog**: 100+ existing agent skeletons, framework comparisons, evaluation tools, tutorials, preserved and improved. ### Tech Stack: Python ### Why It Matters: This handbook provides a production-oriented approach to building, evaluating, securing, and shipping modern LLM agent systems. It covers every layer of the modern agent stack, including identity, memory, skills, MCP, safety, observability, and more. The curated examples and templates make it easy to adopt today.

### Overview: DrivAerNet++: A Comprehensive Dataset for Aerodynamic Car Design DrivAerNet++ is a large-scale multimodal dataset designed to facilitate research and development in the field of aerodynamic car design. This dataset comprises 8,150 diverse car designs modeled with high-fidelity computational fluid dynamics (CFD) simulations, covering configurations such as fastback, notchback, and estateback. The dataset is hosted on Harvard Dataverse and includes a comprehensive set of parameters and configurations to ensure broad applicability and generalization. ### Key Features: - **8,150 Diverse Car Designs**: The dataset encompasses a wide range of car designs, including fastback, notchback, and estateback configurations. - **High-Fidelity Computational Fluid Dynamics Simulations**: Each car design is modeled with high-fidelity CFD simulations to ensure accurate aerodynamic performance. - **Geometric Parameters**: The dataset includes 26 parameters that completely describe each 3D car geometry, allowing for detailed and precise control over design parameters. ### Tech Stack: - **Python**: The primary programming language used to create and manage the dataset. ### Why It Matters: - **Generalization and Accuracy**: The high-fidelity CFD simulations and diverse dataset parameters enable models to generalize better and provide more accurate predictions. - **Comprehensive Coverage**: The dataset covers all conventional car designs, including both ICE and EV vehicles, ensuring broad applicability and generalization. - **Model Validation**: DrivAerNet++ serves as a benchmark for evaluating and validating neural surrogates

### Overview: Lemur is an openly accessible language model optimized for both natural language and coding capabilities, designed to serve as the backbone of versatile language agents. ### Key Features: - **Openness**: Lemur is open-sourced and available on the Hugging Face Hub. - **Foundation Models**: It is pre-trained on a large dataset to provide both strong language understanding and the ability to execute actions. - **Versatility**: Lemur can be used in various applications, including natural language processing and programming. ### Tech Stack: - Python ### Why It Matters: Lemur combines the strengths of both natural language and coding, making it a versatile tool for language agents that can understand and execute tasks effectively. This makes it particularly useful in applications where there is a need for both language understanding and action execution, such as in the field of artificial intelligence and language processing. Repository: OpenLemur/Lemur Primary Language: Python GitHub Stars: 555 GitHub Description: [ICLR 2024] Lemur: Open Foundation Models for Language Agents

Overview: Agent-Kernel is a microkernel multi-agent system framework that leverages large language models (LLMs) for adaptive social simulation. It provides a scalable and dynamic environment for exploring the scaling law of collective intelligence. Key features: - Dynamic addition and removal of LLM-based agents at runtime - Unlimited scalability of agents - Real-time intervention during simulation - Trustworthy verification of agent behaviors and large model outputs - Code reuse across different simulation scenarios Tech stack: Python Why it matters: Agent-Kernel offers a powerful tool for researchers and developers to explore the scaling law of collective intelligence through adaptive social simulation, leveraging the capabilities of large language models to simulate complex social interactions and behaviors. Repository: ZJU-LLMs/Agent-Kernel GitHub stars: 555

Overview: HiRAG is an easy-to-use Graph RAG system designed for efficient knowledge refinement and retrieval, specifically accepted to EMNLP 2025 Findings. It reindexes the knowledge base and integrates hierarchical knowledge, making it suitable for refining knowledge at test time. Key features: - Graph RAG system with hierarchical knowledge integration - Supports re-indexing of knowledge base - Easy-to-use implementation Tech stack: Python Why it matters: HiRAG offers a practical solution for knowledge refinement in natural language processing tasks, making it highly relevant for researchers and developers looking to improve the accuracy and relevance of their language models. Repository: hhy-huang/HiRAG Primary language: Python GitHub stars: 555 GitHub description: [EMNLP'25 findings] An easy-to-use Graph RAG system using hierarchical knowledge.

### Overview: GPT4RoI is a large language model that has been fine-tuned on Region-of-Interest (RoI) data, enabling it to perform single- and multiple-region understanding tasks. This project aims to enhance the capabilities of large language models by leveraging RoI data, making it particularly useful for applications requiring precise region-based analysis. ### Key Features: - **Single-Region Understanding**: GPT4RoI excels at understanding and responding to prompts within a single region. - **Multiple-Region Understanding**: It can handle and respond to instructions across multiple regions, providing comprehensive insights. ### Tech Stack: Python ### Why It Matters: GPT4RoI leverages RoI data to enhance the performance of large language models, making it particularly effective for applications that require precise region-based analysis. This project showcases the potential of fine-tuning models on specific data sets to achieve superior results, demonstrating the power of targeted data in improving model performance. Repository: jshilong/GPT4RoI Primary Language: Python GitHub Stars: 555 GitHub Description: (ECCVW 2025)GPT4RoI: Instruction Tuning Large Language Model on Region-of-Interest

Overview: FinSight AI is an open-source equity research workspace designed for reliable AI agents. It provides evidence-grounded AI research, resilient workflows, versioned reports, and automated quality evaluation. Key features: - Evidence-grounded AI research - Resilient workflows - Versioned reports - Automated quality evaluation Tech stack: Java 17, Spring Boot 3.3.5, PostgreSQL and pgvector Why it matters: FinSight AI offers a focused research workspace, ensuring long-running tasks are recoverable, identical requests are idempotent, reports are bound to data snapshots, and conclusions are inspectable with evidence paths. It is designed for reliable AI agents in the financial sector, providing valuable insights without the risk of automated trading systems.

### Overview: This repository provides an implementation of the MonoDepth algorithm, a state-of-the-art neural network for monocular depth estimation. The code is based on the original work by Clément Godard, Oisin Mac Aodha, and Gabriel J. Brostow, and is optimized for PyTorch. The project aims to create a more lightweight model with improved accuracy, using ResNet50 as the encoder and batch normalization for training stability. The repository includes a dataset for training, a dataloader, and training scripts for both PyTorch and Jupyter notebooks. ### Key features: - PyTorch implementation of MonoDepth - Uses ResNet50 encoder - Batch normalization for training stability - Flexible feature extractor with any version of ResNet from torchvision models zoo - Pretrained models available - Supports both stereo-pair images for training and single images for testing - Dataset: KITTI with 38,237 training samples - Training script for both PyTorch and Jupyter notebooks ### Tech stack: Python ### Why it matters: This implementation provides a lightweight and accurate solution for monocular depth estimation, making it suitable for various applications such as autonomous driving, augmented reality, and robotics. The use of PyTorch and ResNet50 ensures efficient training and accurate results, making it a valuable resource for developers working on depth estimation projects.

Overview: 2txt is an image-to-text tool that leverages the Vercel AI SDK, GPT 5-nano, and Next.js to quickly convert images into text. Key features: - Image-to-text conversion - Fast processing speed - Built with TypeScript, Vercel AI SDK, and GPT 5-nano Tech stack: TypeScript Why it matters: 2txt provides a quick and efficient solution for converting images into text, making it useful for various applications such as customer support, image description, and more.

Overview: 这是一个用C++实现的ASR推理项目，依赖较少，安装简单，推理速度快，可在树莓派4B等ARM平台流畅运行。支持多种模型，包括由Google优化的Transformer模型和开源的Wenetspeech数据集，识别效果优秀，媲美商用ASR软件。项目基于深度学习框架PaddlePaddle或PyTorch实现，语言优势在于C++编译器的优化性能，独立于现有的深度学习框架，依赖较少，执行效率高。 Key features: - 支持多种模型 - 依赖较少，安装简单 - 识别效果优秀 - 可在ARM平台流畅运行 - 语言优势：C++编译器优化性能 - 独立于深度学习框架 - 依赖较少，执行效率高 Tech stack: C++ and PaddlePaddle or PyTorch Why it matters: 该项目实现了4个模型，包括Paraformer、K2RNNT2、Conformer和Conformer Online，识别效果优异，媲美商用ASR软件。项目基于深度学习框架实现，语言优势在于C++编译器的优化性能，独立于现有的深度学习框架，依赖较少，执行效率高，适合在ARM平台流畅运行。

### Overview: This repository is a curated list of artificial intelligence resources, including foundational models, agent frameworks, AI coding tools, and more. It serves as a community-curated resource for those interested in the field of artificial intelligence. ### Key Features: - 🤖 Foundation models: Models like Claude Opus 4.7, GPT-5.5, Gemini 3 Pro, and Grok 4. - 🔓 Open weights: Models like Llama 4, Qwen3, DeepSeek V3.2, and Mistral Large 3. - 💻 AI coding: Tools like Claude Code, Cursor, Codex CLI, and Aider. - 🤝 Agent framework: SDKs for Claude Agent, LangGraph, Anthropic MCP, and Pydantic AI. - 📚 RAG starter: Prototypes like pgvector, Qdrant, and Pinecone. - ⚡ LLM inference: Local models like Ollama, vLLM, and Together AI. - 📊 LLM ops: Observability tools like LangSmith, Helicone, and Braintrust. - 🎓 Free course: Courses like Karpathy's Zero to Hero and HuggingFace Agents Course. - 📰 Newsletter: Newsletters like Latent Space, The Batch, and Import AI. - 📖 Book: Books like AI Engineering and Build a LLM From Scratch. - 🧬 Must-read paper: "Attention Is All You Need (2017)". ### Tech Stack: Python, JavaScript,

### Overview: This repository is a curated collection of resources for applied AI engineering, focusing on the use of Large Language Models (LLMs). It aims to provide developers with essential tools and platforms to deploy and utilize LLMs effectively. ### Key features: - **Serverless Hosting of Private/OS Models**: Various platforms offer private hosting options for models, including Baseten, Modal, and Replicate. - **Model Serving with Lambda and EFS (AWS)**: Demonstrates how to host models using AWS Lambda and Elastic File System. - **BentoML**: A tool for deploying models in a serverless manner. - **LLM Pricing and Limits**: Explains the pricing and limits of serverless compute options like AWS Lambda and Modal. ### Tech stack: Python ### Why it matters: This repository serves as a valuable resource for developers looking to integrate and deploy LLMs efficiently, providing practical examples and comparisons of different hosting options.