Projects

### Overview: Swarm is a Swift-native runtime designed for building type-safe, agent-based AI systems. It supports on-device inference, multi-agent workflows, and memory management, all within the same agent loop, making it ideal for iOS, macOS, and Linux development. ### Key Features: - **Foundation Models**: Supports Apple Foundation Models for on-device inference. - **Composable Workflows**: Allows chaining, parallelizing, and routing agent work. - **Memory Management**: Manages agent memory and guardrails for efficient system operation. - **Streamlined Development**: Simplifies agent development with a lean package and traits for optional surfaces. ### Tech Stack: Swift 6.2 with strict concurrency. ### Why It Matters: Swarm leverages Swift 6.2's strict concurrency, ensuring that values crossing actor boundaries are checked by the compiler. It also provides a seamless on-device inference path for Apple Foundation Models, making it ideal for applications requiring local inference or cloud-based models. The runtime supports workflows, workflows, and workflows, making it a powerful tool for developing complex AI systems in Swift. ### Build a Workflow: Agents can be composed into sequential, parallel, and routed workflows. For example, a researcher agent can be followed by a writer agent, where the output of the researcher becomes the input for the writer. ### Install: Add Swarm to your project using Swift Package Manager. The default package includes core Swarm, Foundation Models, and macros. To enable traits for optional surfaces, add the following lines to your `Package.swift`: ```swift .package( url: "https://

### Overview: Vela is an open-source, privacy-first, local-first AI writing IDE designed for novel writing, web fiction, and creative writing. It integrates LLM capabilities, long-context retrieval (RAG), and automated pipelines for a production-grade novel writing engine. ### Key Features: - **AI-Powered Novel Writing Pipeline**: Deeply integrates LLM capabilities, long-context retrieval (RAG), and automated pipelines for fiction authoring. - 🌍 世界观与设定管理 (Worldbuilding): Custom global world background, core plot axis, character profile archives (dynamic tracking across chapters). - 📋 自动大纲与细纲生成 (Auto Outline): AI generates "structure skeleton → chapter outline → scene/emotion/speed segment requirements". - ✍️ 流式章节正文生成 (Chapter Drafting): Streamlined typing for single chapters, responsive to context and pre-set outlines. - 🔄 AI 智能重写 (Rewrite): Supports selective or full chapter rewriting while maintaining consistency with character and plot. - ✨ 语病与错别字精修 (Refine): AI detects grammar, typos, and logical errors, providing refined suggestions. - 📝 剧情自评审阅 (Review): AI evaluates chapters from a reader/author perspective, highlighting pacing, character arcs, and foreshadowing issues. - 🔁 三重后期管线 (Post-Process): Rewrite → Refine → Review, ensuring each chapter is of high quality. - **Local Knowledge Base**: A

Overview: ICM is a permanent memory system for AI agents that stores and recalls project knowledge using a single SQLite file, eliminating the need for external LLMs. Key features: - Stores project knowledge in a single SQLite file - Automatically extracts facts from tool output without using an embedding model - Shares memory across multiple agents and tools - No external calls to LLMs for storing or recalling memories Tech stack: Rust Why it matters: ICM simplifies memory management for AI agents, reducing the need for external LLMs and improving performance. It's designed to be lightweight and efficient, making it suitable for use in various AI development environments.

Overview: Talkify is a fast, free, and local-first voice dictation tool for macOS, providing on-device transcription and speech recognition capabilities. Key features: - Dictate and translate text - Hands-free session with quick-tapping gesture - Supports second language dictation - Transcribes audio and video files Tech stack: Swift Why it matters: Talkify offers a seamless voice dictation experience on macOS, ensuring privacy and local processing without relying on external services, making it a valuable tool for users seeking a private and efficient dictation solution.

Overview: Solid is a Python framework designed for gradient-free optimization, featuring various algorithms like Genetic Algorithm, Evolutionary Algorithm, Simulated Annealing, Particle Swarm Optimization, Tabu Search, Harmony Search, and Stochastic Hill Climb. Key features: - Implements multiple optimization algorithms without requiring gradient calculations - Facilitates rapid development through ease of use and versatility - Includes detailed documentation Tech stack: Python Why it matters: Solid is essential for anyone looking to learn, modify, or use optimization algorithms without the need for gradient calculations, making it a valuable tool for various applications. Repository: 100/Solid Primary language: Python GitHub stars: 583 GitHub description: 🎯 A comprehensive gradient-free optimization framework written in Python README excerpt: <div align="center"> <img src="logo.png"></img> </div> <br> [](https://travis-ci.org/100/Solid) [](https://github.com/100/Cranium/blob/master/LICENSE) ## *Solid* is a Python framework for gradient-free optimization. #### It contains basic versions of many of the most common [optimization algorithms that do not require the calculation of gradients](https://en.wikipedia.org/wiki/Derivative-free_optimization), and allows for very rapid development using them. #### It's a very versatile library that

### Overview: Opfor is an open-source adversary emulation tool designed to test AI agents, LLM applications, and MCP servers. It simulates attacks from a real attacker's perspective, providing a comprehensive evaluation of AI's response capabilities. ### Key Features: - **Red-Team Scanning**: Emulates attacks on the full AI agent surface, including prompts, tools, MCP servers, memory, and multi-turn reasoning. - **OWASP Top 10 Evaluations**: Generates targeted attacks for OWASP LLM Top 10, OWASP Agentic AI Top 10, OWASP MCP Top 10, OWASP API Security, and EU AI Act bias suites. - **Browser Extension**: A CLI tool that allows non-developers to red-team deployed chatbots without needing to write code or configure environments. - **Command Line Interface (CLI)**: Runs the setup wizard and starts the scan immediately, or saves a configuration for reuse or commit to CI. ### Tech Stack: - TypeScript ### Why It Matters: Opfor is essential for developers and security professionals alike, as it provides a comprehensive and practical way to test AI agents and applications. By simulating real-world attacks, it helps identify vulnerabilities and improve the robustness of AI systems. ### GitHub Stars: 583 ### Repository: KeyValueSoftwareSystems/agent-opfor ### Primary Language: TypeScript ### GitHub Description: Open-source adversary emulation for AI agents and MCP servers. Test your AI like a real attacker would — from your CLI, your IDE, or a browser extension that anyone on your team

### Overview: FaceGate is the world's first Face Authentication enabled MacOS App-locker, providing a strong and secure way to protect your applications. It uses Face, TouchID, or a password to unlock apps, all processed on your Mac without leaving any data. ### Key Features: - **App Locking**: Locks any installed app using FaceUnlock, TouchID, or a password. - **On-Device Face Recognition**: Uses software-based face embedding pipeline running on the Apple Neural Engine. - **Multi-Face Enrollment**: Up to 3 different faces or alternate looks. - **External Camera Support**: Compatible with an external USB webcam. - **Liveness Detection**: Challenges to prevent photo and video spoofing. - **Touch ID**: Seamless integration with macOS Touch ID. - **Password**: Encrypted PIN stored in the macOS Keychain. - **Per-App Session Timers**: Customizable unlock duration per app. - **Lock-on-Sleep**: Automatically locks apps when your Mac sleeps or the screen locks. - **Uninstall Protection**: Prevents casual deletion by making the app bundle immutable. - **Scheduled App-Lock/Unlock**: Time-based auto-lock and auto-unlock windows. - **FaceUnlock Schedule**: Disable face recognition during specific hours. - **Auto-Optimization**: Automatically boosts display brightness and triggers Center Stage. - **Customizable Sensitivity**: Fine-tune face recognition similarity thresholds. - **Menu Bar Control**: Monitor and lock apps directly from the menu bar. - **Secure Operations**: Require authentication to quit the application. - **Emergency Kill Hotkey

### Overview: SER-FIQ, a novel unsupervised method for face image quality assessment, achieves state-of-the-art performance across various databases. ### Key features: - **SER-FIQ (on ArcFace)**: Measuring face image quality based on an arbitrary face recognition model. - **Cross-database Evaluation**: Conducted on three publicly available databases. - **Comparison with State-of-the-Art**: Compared with six state-of-the-art approaches from academia and industry. - **Unsupervised Solution**: Avoids the use of inaccurate quality labels, relying on the deployed face recognition network. - **Stable Performance**: Shows stable performance across all scenarios. - **Training Phase Avoided**: Utilizes the deployed face recognition model for quality assessment, avoiding the training phase. - **Accuracy Improvement**: Outperforms all other approaches in the majority of the investigated scenarios. ### Tech stack: Python ### Why it matters: SER-FIQ provides a novel approach to face image quality assessment that avoids the use of inaccurate quality labels, making it more reliable and accurate compared to previous methods.

Overview: HyDE is a method that enables precise zero-shot dense retrieval without the need for relevance labels, significantly outperforming Contriever across various tasks and languages. Key features: - Precise zero-shot dense retrieval - Significantly outperforms Contriever - Does not require any human labeled relevance judgement Tech stack: Jupyter Notebook Why it matters: HyDE offers a novel approach to zero-shot retrieval that can be applied across different domains and languages, potentially revolutionizing the field of information retrieval. Repository: texttron/hyde Primary language: Jupyter Notebook GitHub stars: 583 GitHub description: HyDE: Precise Zero-Shot Dense Retrieval without Relevance Labels

Overview: This course provides a comprehensive guide for developers looking to build advanced AI systems, focusing on creating a Kubrick AI agent that understands multimodal data from images, video, audio, and text. Key features: - Building a complex multimodal processing pipeline - Creating a video search engine and exposing its functionality to an MCP agent - Developing a production-ready API - Integrating LLMOps principles and best software engineering practices - Implementing custom MCP clients and tools Tech stack: Python, Pixeltable, FastMCP, Opik, Llama 4 Scout, Maverick, Groq, MCP, Prompt versioning Why it matters: This course offers a deep dive into building AI systems that can process and understand a wide range of multimodal data, making it a valuable resource for developers looking to enhance their AI capabilities.

Overview: AI Employe is a browser automation tool that allows developers to teach a human how to interact with a website using GPT-4 Vision. Key features: - Indexing and searching the entire DOM in MeiliSearch for precise element manipulation. - Actions Augmented Generation to prevent GPT from derailing from tasks by recording DOM element changes and embedding them in task prompts. Tech stack: TypeScript Why it matters: AI Employe simplifies browser automation by leveraging GPT-4 Vision for precise and efficient interaction with web elements, addressing common problems in current browser agents.

Overview: This project offers a comprehensive guide for developing a custom inference framework for large language models, specifically focusing on LLaMA2/3 and Qwen2.5. The framework supports Cuda acceleration and Int8 quantization, making it ideal for modern C++ projects. Key features: - Custom inference framework for LLaMA2/3 and Qwen2.5 - Cuda acceleration - Int8 quantization - Supports LLaMA2/3 and Qwen2.5 models - C++ 20 standard code - Excellent project management with CMake and Git - Modern C++ project design and unit testing - Comprehensive documentation and examples Tech stack: C++ Why it matters: This project is ideal for developers looking to gain hands-on experience with large language models, particularly those interested in preparing for school recruitment, autumn recruitment, spring recruitment, and internship opportunities. The project provides a practical, step-by-step guide to building a custom inference framework, ensuring that learners can effectively apply their knowledge to real-world scenarios. Repository: zjhellofss/KuiperLLama GitHub stars: 579

Overview: OlliteRT is an Android application that turns your phone into an OpenAI-compatible LLM inference server, running models on your local GPU/CPU via Google's LiteRT-LM runtime, and serving them as a standard OpenAI-compatible HTTP API. Key features: - Multi-model support - Multimodal & reasoning capabilities - Benchmark built-in - Activity logs - Always-on, low-power configuration - Highly configurable settings - Model & server monitoring - Broad compatibility with Home Assistant, Open WebUI, OpenClaw, Python, and curl Tech stack: Kotlin Why it matters: OlliteRT provides a fully local, private, and open-source solution for running OpenAI-compatible LLMs on Android phones, eliminating the need for cloud services or API keys, making it ideal for developers looking to integrate LLMs into their projects without the complexity of setting up and managing a cloud-based server.

Overview: Memery is a package that enables natural language search over local image datasets, making it easier to find images by describing them with human language. Key features: - Use the browser-based GUI to search images with natural language descriptions. - Use the Python module to integrate image search into existing workflows. - Pipe images through command line tools for direct access. Tech stack: Python, CLIP transformer. Why it matters: Memery simplifies the process of searching through large image datasets by leveraging natural language processing and computer vision, making it accessible and efficient for developers and researchers.

### Overview: Uzi is a powerful CLI tool designed for running multiple coding agents in parallel with Git worktrees. It automates the setup and management of isolated development environments, ensuring efficient collaboration and development workflows. ### Key Features: - **Run Multiple AI Coding Agents in Parallel**: Uzi allows developers to run multiple AI agents simultaneously, enhancing productivity and creativity. - **Automatic Git Worktree Management**: Each agent operates in its own isolated worktree, ensuring clean and isolated development environments. - **Tmux Session Management**: Tmux is used to manage terminal sessions for each agent, providing a seamless development experience. - **Automatic Development Server Setup**: Uzi automatically sets up and manages development servers, ensuring smooth development workflows. - **Real-Time Monitoring**: Real-time monitoring of agent status and code changes helps in tracking progress and debugging issues. - **Automatic Handling of Prompts and Confirmations**: Uzi automatically handles prompts and confirmations, streamlining the development process. - **Easy Checkpoint and Merge**: Agents can be easily checkpointed and merged, ensuring a smooth workflow. ### Tech Stack: Go ### Why It Matters: Uzi simplifies the complex process of parallel coding and development, making it easier for teams to collaborate and enhance their productivity. Its ability to manage multiple agents in isolated worktrees and automate the setup of development environments makes it a valuable tool in the developer's toolkit. Repository: devflowinc/uzi Primary language: Go GitHub stars: 582 GitHub description: CLI for running large numbers of coding agents in parallel with git worktrees

### Overview: MyAgents 是一个免费开源、本地优先的 Agent 桌面端，一站式 Agent 工作台与任务中心。它致力于帮助每个人成为超级个体，通过与智能 Agent 的直接协作，实现高效、智能的工作方式。 ### Key features: - **本地优先的 Agent 工作台与任务系统** - **多种 Agent Runtime**：集成 Claude Agent SDK、DeepSeek Harness 和 Codex，支持多种 AI 订阅登录，配置模型 API。 - **Task 任务系统**：支持定时任务，自动化任务执行，满足条件后唤醒 Agent。 - **Channel IM 通道**：内置飞书、钉钉、Telegram 接入，通过插件扩展更多平台。 - **AgentNet 网络**：同一账号下开放的 Agent 可以跨设备交接工作，使用各自的资料、上下文与执行环境。 - **Space 协作空间**：通过 Issue 和共同目标组织人和 Agent 的协作，共享 Skill 与工具。 ### Tech stack: TypeScript ### Why it matters: MyAgents 通过本地协作和智能 Agent 的无缝集成，为用户提供了一个高效、智能的工作环境。它不仅简化了日常任务的处理，还允许用户通过自己的电脑进行持续研究和开发，无需依赖云端服务。这种工作方式不仅提高了效率，还增强了工作的灵活性和可控性。

Overview: DBreeze is a professional, open-source, multi-paradigm database management system built with C# that supports key-value stores, objects, NoSQL, text search, multi-parameter search, and vector databases. It is ACID-compliant and multi-threaded, making it suitable for servers, desktops, mobiles, and IoT devices. Key features: - Embedded .NET family assembly, platform independent - Multi-threaded, ACID compliant, with deadlock resolution/elimination, parallel reads, and synchronized writes/reads - No fixed scheme for table names (construction and access on the fly) - Tables can reside in mixed locations across different folders and hard drives Tech stack: C# Why it matters: DBreeze is a versatile, open-source database solution that supports various data management paradigms, making it suitable for a wide range of applications. Its ACID-compliant nature ensures reliability, while its embedded architecture ensures platform independence and ease of integration. The ability to construct and access tables dynamically makes it highly flexible and adaptable to different use cases.

Overview: This repository contains the source code for the "Stacked Cross Attention for Image-Text Matching" paper presented at ECCV 2018, focusing on a PyTorch implementation of the SCAN model. Key features: - PyTorch implementation - Source code for "Stacked Cross Attention for Image-Text Matching" paper - Pre-trained models and dataset files for training and testing Tech stack: Python Why it matters: This project provides a practical and efficient way to implement the SCAN model, which is useful for image-text matching tasks.

Overview: This project provides a set of standalone skills for Claude Code and Codex, designed to help developers turn product ideas, code fixes, architecture reviews, and deployments into scoped work with verifiable results. Key features: - `product-discovery-suite`: Evaluates new product opportunities. - `architecture-suite`: Designs system changes with clear boundaries and tradeoffs. - `delivery-planning-suite`: Reviews and plans for deployment. - `implementation-suite`: Fixes, upgrades, simplifies, or speeds up code. - `quality-assurance-suite`: Identifies potential issues. - `delivery-suite`: Publishes or deploys changes. - `operations-suite`: Diagnoses issues or assesses product outcomes. - `skill-maintenance-suite`: Reviews skill boundaries and consistency. Tech stack: Shell Why it matters: These skills help streamline the development process by providing clear, verifiable outcomes for various tasks, ensuring that each step is well-defined and verifiable, thus reducing the risk of errors and improving the overall quality of the project.

Overview: This repository contains a collection of resources for Natural Language Processing (NLP) focused on the task of Event Extraction. Key features: - Pattern Matching - Machine Learning - Deep Learning - Semi-supervised Learning - Unsupervised Learning - Event Coreference - Surveys - Others - Linguistics - Data - Tools and Repos - Other lists Tech stack: Python, NLP libraries, and various NLP tools. Why it matters: This repository provides a comprehensive overview of resources and techniques used in Event Extraction, making it valuable for researchers and developers in the field of NLP. Repository: BaptisteBlouin/EventExtractionPapers GitHub stars: 581 GitHub description: A list of NLP resources focused on event extraction task