Projects

### Overview: Harness is a powerful command center for coding agents and beyond. It allows you to manage and collaborate across multiple machines seamlessly, from your laptop to home servers and even GPUs. With one command center, you can start with code, then explore different disciplines like CAD, circuits, robots, games, and music. ### Key Features: - **Real Terminals:** Every agent runs in its own tmux pane, ensuring seamless scrolling and color support. - **Your CLIs as They Are:** Harness respects your existing CLI configurations without wrapping agents. - **Worktree Per Harness:** Start an agent on its own branch, keeping your working copy clean. - **Key Remapping:** One JSONC file reloaded on save, remapping every key for a comfortable typing experience. - **Local Models:** Run open-weight models on your own machines with Grid, Ollama, MLX-LM, and vLLM. - **Bytes, Not Pixels:** Remote terminals stream text peer to peer, eliminating the need for remote desktops. - **Open Source:** App, CLI, daemon, relay, and device in this repository, ensuring full transparency and flexibility. ### Tech Stack: TypeScript ### Why It Matters: Harness simplifies the process of managing and collaborating with multiple agents and machines, making it easier to work across disciplines and explore different projects efficiently.

### Overview: Transform any face into a video game character, pixel art, claymation, 3D, or toy using the `face-to-many` model. This project allows users to easily convert faces into various digital art styles with just a few clicks. ### Key Features: - Converts faces into 3D, pixel art, video game, and claymation characters. - Compatible with popular platforms like Replicate, ComfyUI, and others. - Includes custom nodes for enhanced functionality. - Supports donations to creators like artificialguybr for further development. ### Tech Stack: Python ### Why It Matters: The `face-to-many` model enables users to quickly transform their faces into engaging digital art, making it a versatile tool for artists and creators looking to add a unique touch to their projects.

### Overview: This project aims to classify emotions on a person's face into one of seven categories using deep convolutional neural networks. The model is trained on the FER-2013 dataset, which consists of 35,887 grayscale, 48x48 sized face images with seven emotions: angry, disgusted, fearful, happy, neutral, sad, and surprised. ### Key features: - Real-time facial emotion detection using deep learning. - Trained on the FER-2013 dataset, published at ICML. - Uses a 4-layer CNN with a test accuracy of 63.2% in 50 epochs. - Compatible with Python 3, OpenCV, and TensorFlow. - Provides basic usage instructions for training and displaying predictions. ### Tech stack: - Python 3 - OpenCV - TensorFlow ### Why it matters: This project is essential for developers interested in real-time emotion detection, particularly in applications like facial recognition systems, security cameras, or user interface enhancements. The ability to classify emotions in real-time can provide valuable insights into human behavior and emotional states, enhancing user experience and security measures.

### Overview: Geti™ is an open-source, end-to-end platform designed for building, training, optimizing, and deploying computer vision models. It supports annotation, fine-tuning, and inference, with a focus on Intel XPU and CUDA optimizations. The platform includes 1,600+ model variants optimized by OpenVINO™ for various hardware types. ### Key Features: - **Model Variants**: 1,600+ model variants for detection, classification, and segmentation. - **Training Engine**: Open-source Python library `getitune` for model training and optimization. - **Deployment**: Supports local and live inference, including annotation and fine-tuning. - **OpenStack**: Built on top of OpenStack, providing a comprehensive solution for computer vision projects. ### Tech Stack: - Python 3.x - OpenVINO™ - Intel XPU - CUDA - Docker ### Why It Matters: Geti™ simplifies the process of building and deploying computer vision models, making it easier for developers to create and deploy models without the need for extensive technical knowledge. The platform's comprehensive feature set and optimized performance make it suitable for a wide range of applications, from small projects to large-scale deployments.

### Overview: A Python extension for the Pi platform that enables Claude Code-style autonomous sub-agents and workflow orchestration, providing parallel execution, live widget UI, fleet view, custom agent types, and more. ### Key features: - **Claude Code look & feel** - Same tool names, calling conventions, and UI patterns. - **Parallel background agents** - Spawn multiple agents that run concurrently with automatic queuing and completion notifications. - **Live widget UI** - Persistent above-editor widget with animated spinners, live tool activity, token counts, and colored status icons. - **FleetView** - Claude Code-style navigable list of main + every running subagent rendered below the editor. - **Conversation viewer** - Select any agent in `/agents` to open a live-scrolling overlay of its full conversation. - **Custom agent types** - Define agents in `.pi/agents/<name>.md` or `.agents/agents/<name>.md` (project) or globally. - **Package-shipped agents** - A pi package can declare its own subagents (and workflow scripts) with `"pi": { "subagents": { "agents": ["./agents"] } }` in its `package.json`. ### Tech stack: TypeScript ### Why it matters: This extension provides a seamless Claude Code experience for Pi, enabling parallel execution, live widget UI, and customizable workflows, making it ideal for developers and users looking to enhance their workflow automation and productivity.

Overview: Infinispan is an open-source in-memory database platform designed for high availability and scalability. Key features: - Key/value data store - Supports Java objects and plain text - Distributes data across elastic clusters for high availability and fault tolerance - Provides flexible deployment options Tech stack: Java Why it matters: Infinispan is crucial for developers looking to build scalable and highly available applications, especially in environments where performance and reliability are paramount.

### Overview: oh-my-agent is a specialized team of coding specialists that enhances the capabilities of your current coding agent. It includes 12 roles, 21 workflows, and 33 skills to streamline your development process. The agent can handle tasks from planning to debugging and even research and content creation. ### Key Features: - **12 Specialist Roles**: PM, architect, backend, frontend, mobile, database, infra engineers, QA reviewer, debugger, refactorer, docs curator, and researcher. - **21 Workflows**: Slash commands for brainstorming, planning, reviewing, debugging, and SCM. - **33 Skills**: Domain knowledge that loads only when needed, such as UI, APIs, schemas, Terraform, CI, security, git, observability. - **Files You Can Open**: Plans, design docs, API contracts, ADRs, bug notes, per-agent results, security evidence, HTML explainers, and recaps. - **Workflow Checks**: Your own `test`, `lint`, or `typecheck` script must pass before the agent stops in long-running workflows. - **Research and Content**: Cited search, market research, papers, slide decks, MP4 videos, images, local voice, PDF/HWP to Markdown. - **One Folder, 14 Runtimes**: One portable `.agents/` directory, written into each runtime's own config files. ### Tech Stack: TypeScript ### Why It Matters: oh-my-agent provides a comprehensive team of specialists to enhance your current coding agent, making your development process more efficient

Overview: RepoBrain is a repository knowledge engine designed to provide grounded answers to codebase questions, integrating with various IDEs and AI tools like Claude Code, Codex, and more. Key features: - Cross-IDE repository knowledge engine - Integrates with multiple AI tools - Tracks and references source files and line numbers for accurate answers Tech stack: Python Why it matters: RepoBrain simplifies understanding unfamiliar repositories, tracing implementations, and asking detailed questions with precise references, making it a valuable tool for developers and AI enthusiasts alike. Repository: study8677/repobrain Primary language: Python GitHub stars: 1351

Overview: DriveLM is a project that introduces a VLM-based baseline approach for joint driving and graph visual question answering. The project is part of the Autonomous Driving Challenge 2024 and serves as a main track. Key features: - DriveLM-Data: Built upon nuScenes and CARLA datasets, DriveLM-Data is a comprehensive dataset for the challenge. - DriveLM-Agent: A VLM-based baseline approach for driving and graph visual question answering. - DriveLM-Data: Contains a complete set of QA pairs covering Perception, Prediction, and Planning QA pairs. - DriveLM-Agent: Demonstrates multi-modal Graph Visual Question Answering (GVQA) benchmark for structured reasoning in Vision Language Models. Tech stack: HTML Why it matters: DriveLM serves as a main track in the Autonomous Driving Challenge 2024, providing a comprehensive dataset and a VLM-based baseline approach for driving and GVQA.

**Overview:** This repository contains a curated list of 167 AI agent harnesses, templates, playbooks, and learning resources. It ranks projects based on their relevance to harness concerns, including environment, orchestration, and lifecycle management. **Key features:** - 🏆 Curated list of AI agent harnesses - 📚 Playbooks for easy setup - 🧑‍💻 Skills checklist for agent setup - 📖 Latest changes every Sunday - 🛠️ Templates and playbooks for harnesses **Tech stack:** Python **Why it matters:** Harness quality is crucial for the reliability and production readiness of agents, ensuring they can handle failures and provide the necessary retry logic, fallbacks, and validation mechanisms. **Repository:** RyanAlberts/best-of-Agent-Harnesses **GitHub stars:** 1322 **GitHub description:** 🏆 Ranked list of 167 AI agent harnesses, plus templates, playbooks, MCP, and learning resources. Rescored weekly.

### Overview: TinyEngram is a research project exploring the DeepSeek Engram architecture and its application in memory injection, focusing on both text and vision tasks. This project aims to demonstrate that Engram-based memory injection outperforms LoRA in terms of parameter efficiency and catastrophic forgetting resistance, with a particular emphasis on extending the architecture to vision tasks like Stable Diffusion. ### Key Features: - **Parameter Efficiency**: Engram-based memory injection demonstrates better parameter efficiency compared to LoRA. - **Catastrophic Forgetting Resistance**: It shows superior resistance to catastrophic forgetting, making it more robust. - **Cross-Modal Experiment**: Successfully injected visual concepts into Stable Diffusion using the Engram architecture. ### Tech Stack: The project is built using Python, leveraging the Qwen-3 model and the Stable Diffusion series for text tasks and vision tasks. ### Why It Matters: TinyEngram provides a compelling demonstration of how Engram-based memory injection can be applied to both text and vision tasks, showcasing improved parameter efficiency and resistance to catastrophic forgetting. This work is significant for researchers and developers looking to enhance the efficiency and robustness of their models, particularly in applications where lightweight and composable concept injection is crucial.

Overview: This book provides an introductory overview of deep generative modeling, covering major classes of models such as mixture models, probabilistic circuits, autoregressive models, flow-based models, latent variable models, GANs, hybrid models, score-based generative models, energy-based models, and large language models. It is designed for students, engineers, and researchers with a modest background in calculus, linear algebra, probability theory, and machine learning. Key features: - Comprehensive coverage of deep generative models - Discussion of Generative AI systems - Introduction to fundamental concepts with examples and code snippets - Designed for beginners and advanced users alike Tech stack: Jupyter Notebook Why it matters: This book serves as an excellent introduction to deep generative modeling, making it accessible to a wide range of readers interested in the field. Repository: jmtomczak/intro_dgm Primary language: Jupyter Notebook GitHub stars: 1347 GitHub description: "Deep Generative Modeling": Introductory Examples README excerpt: # <a href="https://link.springer.com/book/10.1007/978-3-031-64087-2" target="_blank">"Deep Generative Modeling"</a> <img src="figures/1726858979519.jpeg" alt="Deep Generative Modeling"> This first comprehensive book on models behind Generative AI has been thoroughly revised to cover all major classes of deep generative models: mixture models, Probabil

Overview: ShardBrowser Launcher is a free, open-source anti-detect browser launcher designed for web scraping and multi-accounting, featuring engine-level fingerprint spoofing and local API capabilities. Key features: - Engine-level fingerprint spoofing in Chromium 156 - 220+ device profiles bundled - Stable QUIC + WebRTC over SOCKS5 Tech stack: TypeScript Why it matters: It provides a robust solution for bypassing anti-detection measures, making it ideal for web scraping and multi-accounting tasks. Repository: ProxyShard/ShardBrowser Primary language: TypeScript GitHub stars: 1323 GitHub description: Free, open-source anti-detect browser launcher for web scraping and multi-accounting. Patched Chromium 156 with engine-level fingerprint spoofing, a local API, an MCP server and SDKs.

### Overview: Transformer-based models have revolutionized computer vision, enabling advancements in various applications such as autonomous driving, video captioning, and sensor fusion. This repository features recent works and surveys in the field of Transformer-based CV and related technologies. ### Key Features: - **Transformer-based CV Models**: Implementations and analyses of state-of-the-art Transformer-based models for computer vision tasks. - **Sensor Fusion**: Surveys and implementations of Transformer-based methods for sensor fusion in autonomous driving. - **Video Captioning**: Research and implementations of Transformer-based techniques for video captioning. - **Autonomous Driving**: Surveys and implementations of Transformer-based models for autonomous driving applications. ### Tech Stack: Unknown ### Why It Matters: These repositories and surveys highlight the importance of Transformer-based models in advancing computer vision and related technologies. They provide valuable insights into the latest research and practical applications, making them essential for developers looking to stay updated in the field.

### Overview: IMAGDressing is an innovative project that enables the creation of lifelike virtual clothing and scene customization for virtual dressing. The project was accepted by AAAI 2025 and has been widely recognized for its high fidelity and garment consistency. ### Key Features: - **Simple Architecture**: Generates lifelike garments and allows easy scene editing. - **New Task, Metric, and Dataset**: Introduces the virtual dressing (VD) task, designs a comprehensive affinity metric index (CAMI), and releases the IGPair dataset. - **Flexible Plugin Compatibility**: Seamlessly integrates with extension plugins such as IP-Adapter, ControlNet, T2I-Adapter, and AnimateDiff. - **Rapid Customization**: Allows for rapid customization within seconds without the need for additional LoRA training. ### Tech Stack: - Python ### Why It Matters: IMAGDressing is a significant contribution to the field of virtual dressing, offering high-fidelity and flexible customization options that enhance the realism and versatility of virtual clothing.

### Overview: This repository is a curated visual catalog of 155+ Vision-Language Model (VLM/MLLM) architectures, spanning various modalities and applications. It provides primary sources, summaries, and architectural figures for multimodal AI agents, including CLIP, Flamingo, and current models like Mage-VL, InKling, and more. ### Key Features: - **Models:** Over 155 architectures, ordered by release date. - **Release Timeline:** Verified release timeline through July 2026. - **Architectures:** Detailed summaries of each model, including modality alignment, training stages, datasets, and distinctive design choices. - **Citation:** Bibliographic information for each model, including primary sources and citations. ### Tech Stack: Markdown ### Why it Matters: Use this repository to compare multimodal model families, trace architectural ideas over time, or retrieve grounded references for research and AI-agent workflows. The catalog includes a verified release timeline and foundational systems such as CLIP alongside current multimodal reasoning and agentic models. ### Repository: - **Repository:** gokayfem/awesome-vlm-architectures - **GitHub Stars:** 1339 - **GitHub Description:** Curated visual catalog of 155+ vision-language model (VLM/MLLM) architectures: papers, diagrams, training recipes, datasets, and a release timeline for multimodal AI agents. - **Last Reviewed:** August 1, 2026 ### Contents: - **Citation

Overview: 《红楼梦》知识图谱可视化及问答系统，基于知识图谱技术，提供人物关系可视化和问答功能。 Key features: - 人物关系可视化 - 人物关系问答 - 知识图谱构建与查询 Tech stack: Python Why it matters: 该系统利用知识图谱技术，为《红楼梦》的读者提供了一种全新的阅读体验，通过可视化和问答功能，帮助用户更深入地理解小说中的人物关系和情节线索。 Repository: chizhu/KGQA_HLM GitHub stars: 1340 GitHub description: 基于知识图谱的《红楼梦》人物关系可视化及问答系统

Overview: NanoResearch 是一个端到端的自主 AI 科研引擎，它从研究想法到完整论文的每一个环节都实现了自动化，包括生成代码、训练模型、收集实验数据、生成论文配图和最终输出有实验数据支撑的完整 LaTeX 论文。 Key features: - 自动化生成代码 - GPU 集群执行训练 - 实际实验结果生成论文配图 - 自动生成论文配图展示真实实验数据 Tech stack: Python Why it matters: NanoResearch 是一个自主 AI 科研引擎，它不仅能够生成代码，还能将代码提交到 GPU 集群执行训练，收集真实实验结果，生成论文配图，最终输出一篇有实验数据支撑的完整 LaTeX 论文。这使得研究者能够专注于研究本身，而无需担心代码生成和实验结果的收集问题，从而提高了科研效率和质量。

### Overview: This project implements an unorthodox method for fine-tuning the output of Stable Diffusion, known as Cross Attention Control, which allows for more precise control over the generated images without the need for user input. ### Key Features: - **Cross Attention Control**: Fine-tunes the output of Stable Diffusion by modifying the internal attention maps during inference. - **Minimal Performance Penalties**: Maintains performance comparable to other methods without additional training or fine-tuning. - **User-Friendly**: No need for user input, making it accessible for a wide range of users. ### Tech Stack: - **Stable Diffusion**: The model used for image generation. - **Diffusers Library**: Used for implementing the modified prompt-to-prompt editing method. ### Why It Matters: Cross Attention Control provides a more intuitive and precise way to edit images generated by Stable Diffusion, making the process easier and more predictable for users.

### Overview: VeRL-Omni is a comprehensive framework for multimodal reinforcement learning training, designed to optimize the performance of diffusion and omni-modality models. ### Key Features: - **Unified Multimodal Training**: Seamlessly integrates training for image, video, audio, and text data. - **Transfer Learning**: Utilizes pre-trained models to accelerate training times and improve model stability. - **RL Post-Training**: Optimizes the training process for reinforcement learning-based models. - **Easy Setup**: Minimal configuration required for quick deployment and experimentation. - **Stable Training**: Ensures consistent and reliable results across different environments and models. ### Tech Stack: Python ### Why It Matters: VeRL-Omni provides a robust and efficient solution for training multimodal reinforcement learning models, addressing the unique challenges and requirements of these complex systems. Its unified approach and optimized training methods make it a valuable tool for researchers and practitioners in the field of generative AI.