Projects

### Overview: 《御舆：解码 Agent Harness》是一本详细剖析AI代理架构的书籍，涵盖了从对话循环到构建自己的Agent Harness的全过程。本书由LinTsinghua撰写，旨在帮助开发者理解并构建自己的Agent Harness。通过15章和4篇附录，本书结合架构图、设计权衡和动手练习，逐步引导读者深入理解Agent Harness的各个核心部分。 ### Key features: - **Agent Harness 深度剖析**：本书详细解析了Agent Harness的五大设计原则，包括对话循环、工具系统、权限管线和钩子系统。 - **架构图与设计权衡**：通过架构图和设计权衡，帮助读者理解如何在构建过程中权衡不同设计选项。 - **动手练习与工程实践**：提供了构建和优化Agent Harness的实战指南，包括从零构建和性能优化的详细步骤。 - **附录**：包含工具完整清单、功能标志速查表和术语表，方便读者查找和学习。 ### Tech stack: 本书使用Python技术栈进行编写，包括Bun + React/Ink + Zod v4。 ### Why it matters: 理解Agent Harness的构建和运行机制对于任何希望在AI领域进行开发的开发者来说都是至关重要的。本书不仅提供理论知识，还通过实际案例和动手练习帮助读者将理论应用到实践中，从而构建出自己的Agent Harness。

### Overview: ZVT is a modular quant framework designed for developers, offering a comprehensive solution for backtesting, research, and real-time market data handling. ### Key Features: - **Backtesting and Research**: Ideal for developing and testing quantitative trading strategies. - **Real-Time Market Data Handling**: Supports real-time market data and user interaction, suitable for live trading environments. - **Factor Management**: Allows for the management and visualization of trading factors. - **Strategy Execution**: Enables the execution of strategies through both interactive and automated methods. - **Tag System Integration**: Facilitates the integration of AI-driven trading with human intervention through an init tag system. - **API and UI**: Provides a RESTful API for easy integration with other systems and a user-friendly interface for strategy execution and data visualization. ### Tech Stack: Python ### Why It Matters: ZVT offers a robust and flexible platform for quantitative traders, enabling efficient backtesting, real-time market data handling, and seamless integration with AI-driven strategies, making it a valuable tool for developers in the financial industry. ### Repository: zvtvz/zvt ### Primary Language: Python ### GitHub Stars: 4313 ### GitHub Description: modular quant framework.

Overview: SwanLab is an open-source, modern-design AI training tracking and visualization tool that supports both cloud and self-hosted use cases. It integrates with 50+ leading frameworks and seamlessly integrates with your experimental code. Key features: - LightningBoard V2: Enhanced dashboard performance. - LightningBoard V1: Designed for large chart quantities. - Baseline Comparison: Compare experiments with baseline and display difference percentages. - HTML Chart Support: Embed charts in HTML pages. - Parallel Mode: Record metrics in parallel for the same experiment. - Customizable Experiments: Create and replicate experiments. - Experiment Stacking: Place best experiments in a prominent position. - Baseline Monitoring: Monitor experiments against a baseline. - Hardware Monitoring: Monitor hardware performance. - API Integration: Integrate with various AI frameworks. - Skill Integration: Enhance experiments with AI skills. Tech stack: Python Why it matters: SwanLab is essential for AI researchers and data scientists looking to track and visualize their training processes efficiently. Its comprehensive integration with leading frameworks and seamless integration with experimental code make it a valuable tool for modern AI development. Repository: SwanHubX/SwanLab GitHub stars: 4312 GitHub description: ⚡️SwanLab - an open-source, modern-design AI training tracking and visualization tool. Supports Cloud / Self-hosted use. Integrated with PyTorch / Transformers / verl / LLaMA Factory / ms-swift / Ultralytics / MMEngine / Keras etc.

Overview: This repository contains the content used to create the Hugging Face course on Transformers, covering various tasks in natural language processing and beyond. It includes the Hugging Face ecosystem, including Transformers, Datasets, Tokenizers, and Accelerate, as well as the Hugging Face Hub. Key features: - 📚 Transformers - 📂 Datasets - 📄 Tokenizers - 🚀 Accelerate Tech stack: Python Why it matters: This course provides a comprehensive guide to applying Transformers to natural language processing tasks, making it a valuable resource for developers and researchers in the field.

### Overview: NeuralProphet is a simple forecasting package designed for easy learning and interpretability. It combines neural networks with traditional time-series algorithms to provide a human-centered approach to time series forecasting, making it suitable for high-frequency and long-duration data. ### Key Features: - **Ease of Use**: Define, customize, visualize, and evaluate your own forecasting models with a few lines of code. - **Iterative Model Building**: Ideal for iterative human-in-the-loop model building, allowing for quick adjustments and improvements. - **Focus on Interpretability**: Designed for better understanding and customization, it may not be the most accurate out-of-the-box but can be easily adjusted and iterated. ### Tech Stack: NeuralProphet is built on PyTorch, combining neural networks with traditional time-series algorithms. ### Why It Matters: NeuralProphet offers a user-friendly and interpretable approach to time series forecasting, making it ideal for high-frequency and long-duration data. It is particularly useful for iterative model building, allowing for quick adjustments and improvements.

### Overview: DeepFlow is an open-source project designed to provide deep observability for complex cloud-native and AI applications. It leverages eBPF for data collection, supports zero-code instrumentation, and integrates with smart encoding for full-stack correlation and efficient access to all observability data. ### Key Features: - **Universal Map**: Provides a universal map with zero-code by eBPF for production environments, including application services, AI services, and infrastructure services in any language. Supports Wasm plugins for private protocols. - **Distributed Tracing**: Supports distributed tracing powered by eBPF, covering gateways, service meshes, databases, message queues, DNS, NICs, and leaves no blind spots. - **Continuous Profiling**: Collects profiling data with zero-code, plotting flame graphs to locate performance bottlenecks in business functions, library and framework functions, runtime functions, shared library functions, kernel functions, and CUDA functions. - **Seamless Integration**: Integrates with popular observability stacks like Prometheus, OpenTelemetry, SkyWalking, and Pyroscope, and provides SQL, PromQL, and OLTP APIs for data source. - **Performance 10x ClickHouse**: Reduces storage overhead by 10x compared to ClickHouse String or LowCard method, with standardized and pre-encoded meta tags for observability data. ### Tech Stack: Go ### Why it Matters: DeepFlow simplifies observability for cloud-native and AI applications, reducing the burden on developers by eliminating the need to instrument code. It provides comprehensive monitoring and diagnostic capabilities across the entire application

Overview: Ragent AI 是一个面向 Agentic RAG 演进的生产级 Java AI 应用平台，覆盖从文档入库到智能问答的完整链路。它包括混合检索、问题理解、模型与工具、会话记忆、流量保护和知识闭环等功能，旨在提供真实业务场景下的完整工程实践。 Key features: - 混合检索 - 问题理解 - 模型与工具 - 会话记忆 - 流量保护 - 知识闭环 Tech stack: Java Why it matters: Ragent AI 是企业级 Agentic RAG 智能体的完整工程实现，适用于真实业务场景，从 0 到 1 完整工程实现。它覆盖了从文档解析到智能问答的完整链路，提供了一套经过实际锤炼的工程实践，帮助开发者在面试中展示自己的 AI 应用开发能力。

### Overview: Basic Memory is an AI-powered knowledge management system that remembers everything you write, making it accessible to both you and your AI assistants. It works seamlessly with popular AI clients and IDEs, offering a real knowledge graph and progressive tool discovery for a smooth user experience. ### Key Features: - **Local-first.** Plain text on your disk. Forever. - **Two-way.** AI and humans write to the same files; sync keeps them in step. - **A real knowledge graph.** Observations and wikilinks compound into context. - **Semantic search.** Find notes by meaning, not just keywords, with optional cross-encoder reranking for higher-quality vector and hybrid results. - **MCP-native.** Works with every major AI client and IDE. - **Progressive tool discovery.** Every tool is tagged with behavior hints (read-only, destructive, idempotent) so agents pick the right tool on demand — no wasted context trying things to see what they do. - **Cloud, optional.** Sync across devices when you want — never required. ### Tech Stack: Python ### Why it Matters: Basic Memory is designed to be intuitive and efficient, making it perfect for developers and teams looking to streamline their workflow and ensure their knowledge is always accessible and up-to-date.

### Overview: NeuralForecast is a comprehensive library for scalable and user-friendly neural forecasting models, offering a wide range of state-of-the-art algorithms, including popular transformers like `MLP`, `LSTM`, `GRU`, and cutting-edge models like `DeepAR`, `NBEATS`, `NBEATSx`, `NHITS`, `TiDE`, `DeepNPTS`, `TSMixer`, `TSMixerx`, `MLPMultivariate`, `DLinear`, `NLinear`, `TFT`, `Informer`, `AutoFormer`, `FedFormer`, `PatchTST`, `iTransformer`, `StemGNN`, and `TimeLLM`. These models are designed to be fast, accurate, and easy to use, making them ideal for both research and production environments. ### Key Features: - **State-of-the-Art Models**: Offers a collection of over 30 state-of-the-art neural forecasting models, including RNNs, transformers, and more. - **Exogenous Variables and Static Covariates**: Supports the inclusion of exogenous variables and static covariates for improved forecasting accuracy. - **Interpretability**: Provides tools for interpreting trends, seasonality, and exogenous components, enhancing model understanding. - **Probabilistic Forecasting**: Adapts models for probabilistic forecasting with adapters for quantile losses and parametric distributions, offering a range of loss functions. - **Training and Evaluation**: Offers train and evaluation losses with scale-dependent metrics, ensuring accurate model performance. ### Tech Stack: Python ### Why It

### Overview: Perceptual Similarity Metric and Dataset ### Key features: - **LPIPS (Learned Perceptual Image Patch Similarity)**: A metric for comparing image patches that is based on deep neural networks. - **BAPPS (Berkeley-Adobe Perceptual Patch Similarity)**: A dataset containing perceptual similarity patches for various images. ### Tech stack: Python ### Why it matters: LPIPS provides a way to measure image similarity that is more accurate and perceptually meaningful compared to traditional methods. The dataset BAPPS allows for evaluating and training perceptual similarity metrics on real-world images.

### Overview: MNBVC(Massive Never-ending BT Vast Chinese corpus) 是一个超大规模中文语料集，包含主流文化、小众文化甚至火星文的数据。该语料集包括新闻、作文、小说、书籍、杂志、论文、台词、帖子、wiki、古诗、歌词、商品介绍、笑话、糗事、聊天记录等一切形式的纯文本中文数据。MNBVC 数据集的总数据量为 62024GB，目标是达到 253T 数据，目前进度为 24.5%。 ### Key features: - **超大规模语料**：包括主流文化、小众文化甚至火星文的数据。 - **多样数据来源**：涵盖新闻、作文、小说、书籍、杂志、论文、台词、帖子、wiki、古诗、歌词、商品介绍、笑话、糗事、聊天记录等。 - **数据来源信息**：虽然包含数据来源信息，但为了版权问题，不提供索引和分类。 - **数据处理工具**：提供高效的中文编码检测工具、批量转txt工具、数据清洗示例及工具、代码仓库爬虫工具、多模态处理工具等。 - **项目组需求**：包括 OCR转码小组、问答语料小组、语料增强小组、代码语料小组和平行语料小组、古文研究小组、测试组等。 ### Tech stack: Python ### Why it matters: MNBVC 数据集为中文算法圈提供了

Overview: Kelivo is an open-source LLM client built with Flutter for Android, iOS, macOS, Windows, and Linux. It supports connecting to your own API keys or using supported subscriptions, allowing seamless switching between various OpenAI-compatible services without switching apps. Key features: - Cross-platform support - Support for multiple models - Integration with WebDAV or S3-compatible storage for backup - Local storage of conversations, settings, and files - Ability to bind conversations to workspaces and access files and commands - Support for models like OpenAI, Gemini, Claude, DeepSeek, and OpenRouter Tech stack: Flutter Why it matters: Kelivo simplifies the process of using multiple LLMs on different platforms, making it easier to integrate and switch between various models without the need to switch apps. The local storage feature ensures data security and privacy, making it suitable for users who prefer to keep their data on their own devices. Repository: Chevey339/kelivo GitHub stars: 4279 GitHub description: A Flutter LLM Chat Client. Support Mobile & Desktop. GitHub stars: 4279

Overview: This repository demonstrates how to build an **Agentic RAG (Retrieval-Augmented Generation)** system using LangGraph with minimal code. Most RAG tutorials show basic concepts but lack guidance on building modular, agent-driven systems — this project bridges that gap by providing **both learning materials and an extensible architecture**. Key features: - Hierarchical Indexing - Conversation Memory - Query Clarification - Agent Orchestration - Multi-Agent Map-Reduce - Self-Correction - Context Compression - Observability - Evaluation Tech stack: LangGraph Why it matters: This project provides a comprehensive guide for building Agentic RAG systems, making it easier for developers to understand and adapt the architecture. Repository: GiovanniPasq/agentic-rag-for-dummies Primary language: Jupyter Notebook GitHub stars: 4265 GitHub description: A modular Agentic RAG built with LangGraph — learn Retrieval-Augmented Generation Agents in minutes. Installation & Usage: The runnable app is Ollama-first, and it can be adapted to any chat model provider supported by LangChain. Examples are included for Anthropic, OpenAI, and Google. See [Modular Architecture](#modular-architecture) and [Installation & Usage](#installation--usage) to get started.

Overview: Preswald is a Python-based static-site generator designed for building interactive data apps that run in the browser. It bundles complex data workflows, including visualizations, into single files, enabling offline execution and sharing. Key features: - Code-based, file-first approach - Built for computation, using Pyodide and DuckDB - Composable UI with prebuilt components - Reactive engine, only re-running necessary parts - Local execution, no server needed - AI-ready, fully inspectable and modifiable Tech stack: Python, Pyodide, DuckDB, Pandas, Plotly, Matplotlib Why it matters: Preswald simplifies the process of creating data-driven apps, making them portable, fast, and shareable without the need for local dependencies. It's particularly useful for teams working with sensitive data or needing full control over local execution.

Overview: This repository showcases MongoDB's integration into Retrieval-Augmented Generation (RAG) pipelines and AI agents, providing comprehensive examples and sample applications across various industries. Key features: - Retrieval-Augmented Generation (RAG) - AI Agents - Industry-specific use cases Tech stack: Jupyter Notebook Why it matters: This repository offers a practical guide for developers looking to leverage MongoDB's capabilities in AI applications, making it an essential resource for anyone interested in integrating MongoDB with AI technologies.

### Overview: Claude Octopus is an open-source project that allows multiple AI models to review each other's work, ensuring that decisions are made with consensus before they are finalized. It supports twelve external providers and integrates with Claude Code, offering a robust platform for research, design, and coding tasks. ### Key features: - **Multiple Model Reviews:** Claude Octopus allows multiple AI models to review each other's work, ensuring that decisions are made with consensus before they are finalized. - **External Integration:** Supports twelve external provider integrations, including Codex, Antigravity CLI, Copilot, Qwen, Ollama, Perplexity, OpenRouter, OrcaRouter, OpenCode, Cursor CLI, Grok, and Kimi Code. - **Consensus Gate:** A 75% consensus gate flags disagreements before they reach production, ensuring quality and reliability. - **Persistent Memory:** Integrates with Claude Memory, AgentMemory, and Deja-Vu, allowing past decisions, research, and context to survive session boundaries. - **Dark Factory Mode:** Runs a spec and autonomously runs the full pipeline, from research to delivery, with quality gates between phases. - **Persona and Command System:** Supports 31 specialized personas and 54 commands, with 63 reusable workflow modules. - **Zero External Providers Needed:** Adds up to twelve external provider integrations, with each integration becoming available when detected and running only inside an explicit workflow. - **Costs Nothing Extra:** Five providers cost nothing extra when you already have access, including Codex, Antigr

### Overview: Skill Recorder is a desktop app that records and reconstructs your on-screen work session, then converts it into reusable skills or automations for Microsoft's AI agents. ### Key features: - Records your screen and activity locally - Uses GitHub Copilot CLI to reconstruct your actions as an intent and steps - Converts a single run into a Skill or Automation - Supports Skills and Automations - Reuses native tools like `gh` CLI or `web_fetch` - Provides an easy-to-use interface for recording, analyzing, and creating skills ### Tech stack: TypeScript ### Why it matters: Skill Recorder simplifies the process of creating reusable skills and automations for Microsoft's AI agents, making it easier to automate tasks and improve productivity.

Overview: This comprehensive repository is designed to guide developers through the entire journey of becoming proficient in Artificial Intelligence, Machine Learning, Generative AI, Deep Learning, Data Science, Natural Language Processing, Reinforcement Learning, and more. Key features: - Free AI and ML resources - Step-by-step learning modules - Recommended content for each module - YouTube video tutorials - Python installation guide - Math and programming fundamentals Tech stack: Python, Visual Studio Code, Pip, Common Python Libraries Why it matters: This repository provides a structured and accessible way for developers to learn and grow in the rapidly evolving field of AI and ML, ensuring they are equipped with the necessary skills and knowledge to tackle real-world projects and challenges. Repository: aadi1011/AI-ML-Roadmap-from-scratch GitHub stars: 4251

### Overview: Excel MCP Server is an open-source Model Context Protocol (MCP) server designed to enable AI clients like Claude, Cursor, VS Code Copilot, and Codex to create, read, and edit Excel files without the need for Microsoft Excel. It leverages the `openpyxl` library and supports charts, PivotTables, formulas, and dynamic arrays, providing a seamless experience for AI-driven data manipulation. ### Key Features: - **Charts and PivotTables**: Create Excel charts with AI, including advanced features like combo charts, secondary axes, trendlines, and error bars. Real PivotTables that Excel refreshes, including slicers, timelines, and tables. - **Formulas and Dynamic Arrays**: Formulas are calculated by a built-in engine, allowing `read_range` to return values for files that were never opened in Excel. Dynamic arrays are stored as Excel stores them and spill into the cells next to them. - **Data, Tables, and Rules**: Read and write cells, dates, hyperlinks, tables with totals rows and calculated columns, conditional formatting, data validation, sorting, filtering, and more. Supports large files by reading them in chunks and updating references. - **Large Files**: Reads and edits files efficiently by parsing only the rows that are touched, ensuring memory usage remains flat on large worksheets. ### Tech Stack: - Python - `openpyxl` ### Why It Matters: Excel MCP Server makes it possible for AI clients to create, read, and edit Excel files seamlessly, without the need for Microsoft Excel. It simplifies the process of data

Overview: Kubeflow Pipelines is a machine learning (ML) toolkit designed for deploying ML workflows on Kubernetes, simplifying the orchestration of end-to-end ML pipelines. Key features: - Reusable end-to-end ML workflows built using the Kubeflow Pipelines SDK - End-to-end orchestration for ML pipeline deployment - Easy experimentation with numerous ideas and techniques - Easy re-use of components and pipelines for quick solution creation Tech stack: Go Why it matters: Kubeflow Pipelines simplifies the deployment and orchestration of machine learning workflows on Kubernetes, making it accessible and scalable for developers and data scientists. Repository: kubeflow/pipelines GitHub stars: 4248