Projects

Overview: Harvester is a 2D space exploration and mining game built with Godot and our AI framework, featuring procedurally generated maps, AI-controlled pirates, and player-controlled upgrades. Key features: - Procedural map generation - AI-controlled pirates - Player-controlled upgrades - Godot game engine Tech stack: GDScript Why it matters: This game showcases the power of Godot and our AI framework, providing a fun and educational experience for developers. Repository: gdquest-demos/godot-2d-space-game GitHub stars: 1111 GitHub description: A 2D space exploration and mining game made with Godot and our AI framework

### Overview: This runbook provides a practical guide for running OpenClaw agents without burning money, exposing your gateway, or trusting random automation. It is opinionated, focusing on models, access, memory, automation, and guardrails, and is written from the post-honeymoon phase. ### Key features: - **Model Routing**: Configured model routing for optimal performance. - **Memory Management**: Efficient memory management strategies. - **Automation**: Built-in automation for tasks and agents. - **Skills**: Skeptical approach to third-party skills. - **Tailscale Access**: Tailscale-first access for remote control and dashboard access. - **Security**: Security baseline and audit workflow. - **Examples**: Templates and references for various use cases. ### Tech stack: Astro ### Why it matters: This runbook is a practical guide for running OpenClaw agents without the need to burn money, expose your gateway, or trust random automation. It is opinionated, focusing on models, access, memory, automation, and guardrails, and is written from the post-honeymoon phase. ### Repository: digitalknk/openclaw-runbook ### GitHub stars: 1110 ### GitHub description: Unofficial OpenClaw runbook for running agents day to day without burning money, exposing your gateway, or trusting random automation.

### Overview: 本项目是一个基于GPT2模型的中文新闻标题生成系统，通过详细的注释和详细的代码结构，帮助开发者快速上手并实现新闻标题的自动生成功能。 ### Key features: 1. **详细注释的GPT2模型**：使用HuggingFace的transformers库实现GPT2模型的编写、训练和测试。 2. **数据集整理**：从多个公开数据集收集新闻数据，并进行了清洗和整理，构建了一个较完善的中文摘要数据集。 3. **Web服务部署**：通过Flask框架搭建了一个Web服务，用户可以通过页面可视化地体验新闻标题生成效果。 4. **训练和测试**：使用5轮训练，随机初始化参数，训练效果一般，但为后续的模型优化提供了基础。 5. **数据集**：包含清华新闻数据、搜狗新闻数据、nlpcc2017摘要数据、csl摘要数据、教育培训行业摘要数据、lcsts摘要数据、神策杯2018摘要数据、万方摘要数据、微信公众号摘要数据、微博数据等。 6. **模型参数**：详细见config/config.json文件。 ### Tech stack: Python 使用Python语言进行项目开发，涉及HuggingFace的transformers库、Flask框架等。 ### Why it matters: 本项目为开发者提供了一个详细的GPT2模型训练、测试及部署的流程，帮助他们快速上手并实现新闻标题的自动生成功能。通过详细的注释和数据集

### Overview: This repository is a curated collection of resources focused on **controllable generation** with **text-to-image diffusion models**. It serves as a survey paper and a repository for various techniques and applications in this field. ### Key features: - **Personalization**: Techniques for personalizing text-to-image diffusion models. - **Subject-Driven Generation**: Generating images based on specific subjects. - **Person-Driven Generation**: Personalizing images based on specific individuals. - **Style-Driven Generation**: Generating images with specific styles. - **Interaction-Driven Generation**: Generating images based on interactions. - **Image-Driven Generation**: Generating images from specific images. - **Distribution-Driven Generation**: Generating images from specific distributions. - **Spatial Control**: Techniques for controlling the spatial aspects of generated images. - **Advanced Text-Conditioned Generation**: Enhancing the text-conditioned generation of diffusion models. - **In-Context Generation**: Generating images in the context of specific prompts. - **Brain-Guided Generation**: Generating images based on brain activity. - **Sound-Guided Generation**: Generating images based on audio inputs. - **Text Rendering**: Techniques for rendering text in generated images. ### Tech stack: The repository is built using Markdown and GitHub Pages. ### Why it matters: This collection provides a comprehensive overview of the latest advancements in controllable text-to-image diffusion models, making it valuable for researchers, developers, and enthusiasts alike.

### Overview: WhisperBoard is an open-source iOS app designed to enhance quality voice transcription on mobile devices, utilizing OpenAI's Whisper model for unparalleled accuracy. ### Key Features: - **Simplicity at Your Fingertips**: Start recording with a single tap and play back your audio with ease. - **Transcription Magic**: Transcribe your audio with cutting-edge technology powered by OpenAI's Whisper model. - **Audio File Mastery**: Import existing audio files or export new ones for seamless sharing and editing. - **Mic Check**: Choose your preferred microphone to ensure the best sound quality for your recordings. - **Model Flexibility**: Browse and download any Whisper model directly from the app to tailor your transcription experience. ### Tech Stack: Swift ### Why it Matters: WhisperBoard is revolutionizing the accessibility of voice transcription on mobile devices, making it more accessible and efficient for users.

### Overview: Conformal Prediction is a rigorous uncertainty quantification method that provides confidence intervals for predictions, ensuring that the intervals contain the true values with a specified probability. This lightweight implementation is designed to be easily integrated into any real-world machine learning task, making it accessible to developers and researchers alike. ### Key Features: - **Conformal Prediction Implementation**: A practical and efficient implementation of conformal prediction. - **Real Data Applications**: Applies conformal prediction to real-world datasets, including ImageNet classification, medical expenditure regression, multi-label image classification, and outlier detection. - **Automatic Data and Model Management**: Automatically downloads necessary model outputs and data for each notebook, allowing users to develop and test conformal prediction methods without needing to run the original model or download the original data. ### Tech Stack: - **Python**: Utilizes Jupyter Notebook for interactive data analysis and machine learning. ### Why It Matters: Conformal Prediction is crucial for providing reliable and robust predictions, especially in scenarios where uncertainty is critical, such as medical diagnosis, financial forecasting, and machine learning model evaluation. By ensuring that prediction intervals contain the true values with a specified probability, conformal prediction helps in making more informed and trustworthy decisions.

### Overview: Loki Mode is an autonomous software factory that turns tasks into verified pull requests with signed receipts. It runs on your machine using your own keys, making it accessible and secure. ### Key features: - **Task to Pull Request Conversion**: Converts any task or issue into a verified pull request with a signed receipt. - **Offline Verification**: Allows you to check the pull request without a provider or key. - **CLI Interface**: Provides a command-line interface for easy execution. - **Provider Options**: Supports multiple coding CLIs, including Claude, Codex, and OpenCode. ### Tech stack: - **Shell**: The primary language used for the project. ### Why it matters: Loki Mode streamlines the software development process by automating the verification and integration of changes, ensuring that only verified and secure changes are committed to the codebase. ### Repository: The repository is maintained by **asklokesh** and has been starred by **1104** users, indicating a high level of interest and adoption. ### Install: ```bash npm install -g loki-mode ``` ### Usage: - **Start**: Runs the Loki engine and starts a new project. - **Edit**: Allows for small changes to be made and verified. - **Chat**: Provides an interactive alternative to the CLI. - **Route**: Provides a way to inspect the route without running the task. ### Requirements: - **Claude Code**: The default provider, required for the CLI. - **ANTHROPIC_API_KEY** or **OPENAI_API_KEY**: Optional, for more

Overview: WhisperWriter is a small speech-to-text app that uses OpenAI's Whisper model to auto-transcribe recordings from a user's microphone to the active window. Key features: - Continuous recording mode - Voice activity detection mode - Press-to-toggle recording mode - Hold-to-record mode - Configuration options for keyboard shortcut and recording mode - Transcription can be done locally or through an API Tech stack: Python Why it matters: WhisperWriter is a useful tool for those who need to transcribe audio recordings quickly and easily, and it provides a simple and intuitive interface for doing so.

Overview: Pydantic Deep Agents is an open-source, self-hosted terminal AI assistant and Python framework for building your own agents. It allows you to plan, edit files, run commands, search the web, remember across sessions, spawn sub-agents, and connect to MCP servers using any model. Key features: - Terminal AI Assistant - Python Framework - Self-hosted - Type-safe - Open-source (Apache-2.0) - Built on Pydantic AI - Supports any model Tech stack: Pydantic AI Why it matters: Pydantic Deep Agents provides a powerful and versatile tool for developers and AI enthusiasts alike, offering a self-hosted alternative to Claude Code and a straightforward way to build your own AI agents using a single function call.

### Overview: Rex is an open-source quadruped robot designed to explore reinforcement learning and OpenAI Gym environments. The project aims to train the robot using reinforcement learning techniques to perform domestic and generic tasks in simulations before transferring the knowledge to the real robot without manual tuning. ### Key features: - OpenAI Gym environments for training the robot. - A web URDF visualizer for Rex robot models. - A CLI application to bootstrap and control the Rex robot. - Scripts for batch training, policy reproduction, and single training sessions. - Pre-trained agents for playing learned control policies. ### Tech stack: Python ### Why it matters: This project provides a comprehensive framework for training and deploying reinforcement learning agents on an open-source quadruped robot, making it accessible and reusable for researchers and developers interested in robotics and reinforcement learning.

Overview: AI-Programmer is an experiment using artificial intelligence and genetic algorithms to automatically generate programs. The project aims to create a computer program that can autonomously write software programs to solve specific computational tasks. Key features: - Successfully created programs including "hello world", "hello <name>", addition, subtraction, reversing a string, fibonnaci sequence, and more. - Uses an esoteric programming language for generating software programs. - The underlying programming language consists of only 8 instructions, making it Turing complete. - The AI is capable of solving any computational problem theoretically, but this experiment focuses on outputting simple strings. Tech stack: C# Why it matters: This project demonstrates the potential of using AI and genetic algorithms to automate the creation of software programs, potentially leading to more efficient and tailored solutions in the future.

Overview: JamAI Base is an open-source RAG backend platform that integrates an embedded SQLite database and LanceDB vector database, providing built-in LLM, vector embeddings, and reranker orchestration. It features a spreadsheet-like UI and a simple REST API, making it ideal for collaborative AI development. Key features: - Embedded SQLite database and LanceDB vector database - Managed memory and RAG capabilities - Built-in LLM, vector embeddings, and reranker orchestration - Intuitive spreadsheet-like UI - Simple REST API Tech stack: Python Why it matters: JamAI Base simplifies the development and management of AI applications, making it easy to collaborate and iterate on AI projects.

### Overview: Backburner is a project that leverages your iPhone to speed up prompt reading and enhance context for Qwen3.8-27B on a Mac. This setup allows for faster prefill and more context, making the agent's performance more efficient. ### Key Features: - **Faster Prefill (up to 64k context):** The Mac runs layers 1-40, while the iPhone runs 41-64, pipelined. This results in 29-44% faster prefill at 16k-48k. - **More Context:** The Mac holds 64k tokens of context, while the iPhone holds the oldest part past that. The iPhone's GPU and Neural Engine compute attention over the old keys while the Mac runs all 64 layers. - **Same Answers:** Greedy output is token-identical with and without the phone. ### Tech Stack: The project utilizes Python, llama.cpp fork, and its own Mac kernels (SME2, Metal fusions, DFlash2 speculative decoding). ### Why It Matters: By leveraging the iPhone to speed up prompt reading and enhance context, Backburner allows for more efficient performance, making the agent's performance more efficient and faster. ### Repository: StayLameBro/backburner ### Primary Language: Python ### GitHub Stars: 1035 ### GitHub Description: Your iPhone helps your Mac run a 27B model: faster prompt reading and more context over a USB-C cable.

### Overview: 本项目是“从 NLP 到 LLM 的算法全栈教程”，旨在为开发者提供从理论到实战的全面指南。通过系统回顾 NLP 的核心技术，从经典算法到大模型架构，帮助读者构建坚实的工程化思维能力。 ### Key features: - **理论基础**：深入讲解分词、词向量、RNN/LSTM、Transformer等经典算法。 - **核心架构**：剖析 Transformer 架构，理解 Attention 机制，掌握 Encoder-Decoder 架构。 - **里程碑模型**：全面覆盖 BERT、GPT、T5 等大模型的设计与应用。 - **实战项目**：从零手搓 Llama2 模型，掌握参数高效微调、RLHF 等前沿技术。 - **工程化实践**：涵盖模型量化、推理加速、Docker 容器化及服务部署的全流程实战。 - **安全与多模态**：探索模型安全挑战、伦理问题，以及图文多模态模型的前沿技术。 ### Tech stack: Python ### Why it matters: 掌握大语言模型已成为 AI 工程师的必备技能。本项目通过系统性地回顾 NLP 的核心技术，帮助开发者构建坚实的工程化思维能力，提升在实际项目中的应用能力。

Overview: This project implements the QuickDraw game, a popular online game developed by Google, allowing users to draw with a webcam or on a canvas. It includes both a camera-based and a canvas-based application, with the ability to train a model using a dataset of 20 categories. Key features: - Camera-based drawing app - Canvas-based drawing app - Model training with a dataset of 20 categories Tech stack: Python Why it matters: This project provides a practical implementation of a popular Google game, demonstrating how to develop and train a model for image classification tasks using a dataset of 20 categories.

Overview: The `octo-server` is the core backend of OCTO, powering REST and WebSocket APIs, orchestrating Lobster AI agents, and managing the WuKongIM real-time messaging system. Key features: - REST + WebSocket APIs - Lobster AI agent orchestration - WuKongIM real-time messaging control plane Tech stack: Go Why it matters: The `octo-server` serves as the central anchor for the entire OCTO platform, integrating clients, adapters, matter, and admin functionalities. It handles routing, session management, and tool-call execution for OpenClaw-powered digital doubles, ensuring they are treated as first-class conversation participants. The server also supports pluggable storage and IM systems, making it highly flexible and customizable.

Overview: This repository is a curated list of open-source projects and applications integrated with Hugging Face libraries, showcasing their state-of-the-art capabilities in natural language processing and related fields. Key features: - 🤗 Official Libraries: A collection of popular Hugging Face libraries including transformers, datasets, tokenizers, knockknock, accelerate, autonlp, nn_pruning, huggingface_hub, tune, and more. - 👩‍🏫 Tutorials: Comprehensive guides on using Hugging Face toolkits, provided by the official course series and tutorials by @nielsrogge. - 🧰 NLP Toolkits: A variety of NLP toolkits built upon Transformers, such as AllenNLP, Graph4NLP, Lightning Transformers, Adapter Transformers, Obsei, Trapper, and Flair. - 🥡 Text Representation: Sentence embedding techniques like Sentence Transformers, WhiteningBERT, SimCSE, and DensePhrases. - ⚙️ Inference Engines: High-performance inference engines for transformers, including TurboTransformers and FasterTransformer. Tech stack: Python, JAX, PyTorch, TensorFlow, JavaScript, C++, and more. Why it matters: This repository serves as a valuable resource for developers and researchers interested in leveraging Hugging Face's cutting-edge NLP tools and techniques, providing a comprehensive overview of the latest developments in the field.

### Overview: This JavaScript/WebGL library enables real-time face tracking and expression detection in the browser, with both SVG and Three.js demos included. It utilizes the computing power of the user's GPU for smooth and accurate results, making it suitable for mobile-friendly applications. ### Key features: - Real-time face tracking and expression detection - 11 facial expressions detection - Face rotation around the 3 axes - Robust to lighting conditions - Mobile-friendly - Examples provided using both SVG and THREE.js ### Tech stack: JavaScript ### Why it matters: This library simplifies the process of adding animated emoticons to web applications, leveraging the power of the GPU for real-time performance and accuracy. It's particularly useful for developers looking to integrate facial expressions into their projects without the need for complex backend setups.

### Overview: TabTransformer is an attention-based model designed for tabular data, implementing the TabTransformer architecture from the paper "TabTransformer: Tabular Data Modeling Using Contextual Embeddings". This implementation is built on Pytorch and aims to improve upon the performance of Gradient Boosting Decision Trees (GBDT) on tabular datasets. ### Key Features: - **TabTransformer Architecture**: Implements the TabTransformer model for tabular data classification. - **Attention Mechanism**: Utilizes attention mechanisms to capture the relationships between different categories and continuous features. - **Performance Comparison**: Competes with Gradient Boosting Decision Trees (GBDT) on real-world tabular datasets, achieving results within a hair's breadth of GBDT's performance. ### Tech Stack: - **Python**: The primary programming language used for implementing TabTransformer. - **Pytorch**: The deep learning framework used to build and train the TabTransformer model. ### Why It Matters: TabTransformer represents a significant advancement in the field of tabular data modeling, offering a more efficient and effective approach compared to traditional methods like Gradient Boosting Decision Trees (GBDT). The attention mechanism allows for better handling of categorical and continuous features, potentially leading to improved model performance on various tabular datasets. ### Repository: The repository is named `tab-transformer-pytorch` and is hosted on GitHub. It has received 1099 GitHub stars, indicating its popularity and utility in the community. ### Install: To use TabTransformer, install the library via pip: ```bash $ pip install tab-transformer-pytorch

Overview: CogView4, CogView3-Plus, and CogView3 are advanced multimodal generation models designed for various applications, including text-to-image and image-to-text conversion. Key features: - **CogView4**: A powerful multimodal generation model that supports Chinese text-to-image conversion. - **CogView3-Plus-3B**: A text-to-image model with 3B parameters, supporting Chinese text-to-image generation. - **CogView3**: A text-to-image model based on cascading diffusion, using a relay diffusion framework. Tech stack: The models are based on Hugging Face's Diffusers framework. Why it matters: These models are highly efficient and capable of generating high-quality images from text inputs, making them valuable tools for various applications in the field of AI and computer vision.