Projects

Overview: This book aims to demystify the art of scaling Large Language Models (LLMs) on TPUs, explaining how TPUs work, how LLMs run at scale, and how to pick parallelism schemes that avoid communication bottlenecks. Key features: - Demystifying TPUs and LLM scaling - Explanation of parallelism schemes for training and inference - Avoiding communication bottlenecks during training and inference Tech stack: Ruby, ImageMagick, Jupyter Why it matters: This book provides practical insights into scaling LLMs on TPUs, making it valuable for developers and researchers looking to optimize model performance on hardware. Repository: jax-ml/scaling-book Primary language: SCSS GitHub stars: 1487 GitHub description: Home for "How To Scale Your Model", a short blog-style textbook about scaling LLMs on TPUs GitHub Actions: GitHub Pages deployment Running Locally: Ruby, ImageMagick, Jupyter installation Generating a Single Document: Python script to convert chapters to markdown and combine them Contributing: Giscus discussion, GitHub PRs, CLA signing Citation: Google DeepMind, 2025, BibTeX citation available

Overview: KoNLPy is a Python package designed for natural language processing in the Korean language. Key features: - Comprehensive Korean language processing capabilities - Extensive support for Korean-specific NLP tasks - Easy integration with other Python libraries and frameworks Tech stack: Python Why it matters: KoNLPy provides a robust and versatile solution for Korean NLP tasks, making it an essential tool for researchers, developers, and language enthusiasts working with the Korean language.

Overview: DOX is a self-documenting AGENTS.md framework that ensures precise project context for AI agents. Key features: - Traverses the project's AGENTS.md hierarchy - Provides local guidelines for precise edits - Keeps the docs current with meaningful changes Tech stack: AGENTS.md Why it matters: Simplifies project management by reducing guesswork and drift, ensuring the project's documentation remains up-to-date and contextually relevant.

### Overview: This repository is a curated list of research papers, projects, and agent skills focused on using Large Language Models (LLMs) for software vulnerability detection and discovery. The papers cover various aspects such as repository-level, agentic, and smart-contract detection, and include datasets, benchmarks, and surveys. ### Key Features: - **VulnGym**: A benchmarking repository for coding agents for repository-level vulnerability detection. - **VulTriage**: Improves LLM-based vulnerability detection by augmenting triple-path context. - **Synthesizing Multi-Agent Harnesses**: Enhances vulnerability discovery through multi-agent harnesses. - **QRS**: A rule-synthesizing neuro-symbolic triad for autonomous vulnerability discovery. - **Seclens**: Evaluates LLMs for security vulnerability detection with role-specific evaluation. - **Do Fine-Tuned LLMs Understand Vulnerabilities?**: Investigates semantic traps in LLMs. - **Sifting the Noise**: Compares LLM agents in vulnerability false positive filtering. - **AgenticSCR**: An autonomous agentic secure code review for immature vulnerability detection. - **LLM-based Vulnerability Detection at Project Scale**: An empirical study on vulnerability detection at project scale. - **MulVul**: A retrieval-augmented multi-agent code vulnerability detection system. - **VULPO**: Context-aware vulnerability detection using on-policy LLM optimization. - **VulInstruct**: Teaches LLMs root-cause reasoning for vulnerability detection. - **From Large to Mammoth**: A comparative evaluation of large language models

Overview: PeterCat 是一个专为社区维护者和开发者设计的智能答疑机器人解决方案，提供对话式配置系统、自托管部署方案和便捷的应用 SDK，帮助用户快速创建和集成智能答疑机器人。 Key features: - 自动创建机器人 - 知识自动入库 - 多平台集成 - 项目信息查询 Tech stack: TypeScript Why it matters: PeterCat 通过简化机器人创建流程、自动知识入库和多平台集成，为用户提供了一站式的智能答疑解决方案，有效提升社区和开发者的技术支持效率。

Overview: Dust is a cutting-edge AI agent platform designed to streamline your workflow by automating repetitive tasks and providing personalized insights. Key features: - Customizable AI models for tailored solutions - Integration with various development tools and platforms - Real-time updates and notifications for seamless collaboration Tech stack: TypeScript Why it matters: Dust simplifies complex AI integration, enhancing productivity and efficiency across different development environments, making it a valuable tool for developers looking to accelerate their projects.

Overview: This repository is a curated list of AI-powered web search software, focusing on the intersection of Large Language Models (LLMs) and web search capabilities. The software supports various use cases such as web search with LLM summarization and follow-up questions, LLM chat with web search support via function calling, and agent-driven web research for report generation. Key features: - Web search with LLM summarization and follow-up questions - LLM chat with web search support via function calling - Agent-driven web research for report generation Tech stack: HTML Why it matters: This list provides a comprehensive overview of AI-powered web search software, making it easier for developers to find and evaluate the best options for their projects. Repository: felladrin/awesome-ai-web-search Primary language: HTML GitHub stars: 1485 GitHub description: List of software that allows searching the web with the assistance of AI: https://hf.co/spaces/felladrin/awesome-ai-web-search

Overview: Datalevin is a simple, fast, and versatile Datalog database designed for easy querying and efficient storage. It aims to bridge the gap between declarative Datalog and the practical needs of modern databases. Key features: - Datalog query language for efficient data retrieval - Robust ACID transaction support - Lightweight and fast performance - Built-in support for WAL and asynchronous transactions Tech stack: Clojure Why it matters: Datalevin simplifies Datalog querying by providing a simple, fast, and versatile database engine that caters to both beginners and advanced users. Its ACID compliance and efficient storage capabilities make it suitable for a wide range of applications, from small projects to large-scale data management.

Overview: This repository contains the Human vs. ChatGPT comparison corpus, HC3, and various detectors for content detection. It serves as a resource for evaluating and comparing the performance of ChatGPT with human experts. Key features: - Human vs. ChatGPT comparison corpus (HC3) - Detectors for content detection Tech stack: Python Why it matters: This repository provides a valuable resource for researchers and developers to evaluate the performance of ChatGPT against human experts, contributing to the field of AI and natural language processing.

### Overview: Phantom is an AI co-worker with its own computer, built on the Claude Agent SDK. It installs software, builds infrastructure, and learns from interactions, making it more effective over time. Phantoms have created a variety of tools, including an analytics platform, a Discord integration, and a secure credential collection system. ### Key features: - **Built an analytics platform from scratch**: Installed ClickHouse on its own VM, processed 28.7 million rows of Hacker News data, and created an analytics dashboard. - **Extended itself with a channel it was never built with**: Successfully integrated Discord on the fly when asked. - **Started monitoring its own infrastructure**: Built a system to monitor service health and Docker container status. ### Tech stack: TypeScript ### Why it matters: Phantom demonstrates the potential of AI agents that can adapt to their environment and build tools without being explicitly programmed. This approach could revolutionize how AI is integrated into everyday applications, making them more powerful and useful.

### Overview: LightAutoML is a fast and customizable framework for automatic machine learning model creation, providing tabular, time series, image, and text data support. ### Key features: - Tabular data preprocessing and model creation - Time series forecasting using LSTM - Image classification using CNNs - Text classification using transformers - Customizable pipelines with ready-to-use blocks - Easy integration with existing ML models ### Tech stack: Python ### Why it matters: LightAutoML simplifies the process of building machine learning models, making it accessible for developers and researchers alike.

Overview: ChainerCV is a library designed for deep learning tasks in computer vision, providing tools to train and run neural networks for various vision-related tasks such as image classification, object detection, semantic segmentation, and instance segmentation. It aims to simplify the process of developing computer vision models by offering a consistent and easy-to-use interface. Key features: - Comprehensive support for image classification, including ResNet, SENet, and VGG models. - Object detection capabilities with support for Faster R-CNN, FPN, SSD, YOLO, and Light-Head R-CNN. - Semantic segmentation tools for models like SegNet, PSPNet, and DeepLab. - Instance segmentation tools for FCIS and Mask R-CNN. Tech stack: ChainerCV is built on top of Chainer, a deep learning framework for Python, and it utilizes Python for its core functionalities. Why it matters: ChainerCV simplifies the development of computer vision models by providing a cohesive and user-friendly interface, ensuring reproducibility through well-designed training scripts, and enabling the composition of various tools for efficient model development and deployment. Repository: The repository is hosted on GitHub under the name "chainer/chainercv", with 1481 GitHub stars. It is part of the Chainer project, which is a community-driven initiative for deep learning in Python. Primary language: The primary language used in ChainerCV is Python, making it accessible and easy to integrate into existing Python-based projects. Installation: ChainerCV can be installed using pip, and it requires dependencies such as Chainer,

Overview: This project provides a comprehensive set of professional agent skills for Claude, ChatGPT, Gemini, Cursor, and Codex, including 1255 skills, including PRDs, postmortems, leases, medical bills, layoffs, go-bags, and more. The skills are available in plain markdown format and are licensed under MIT. The skills are also included in Anthropic's plugin directory, and a typed decision layer is provided to ensure safety before deployment. Key features: - 1255 professional agent skills - Available in plain markdown format - MIT license - Included in Anthropic's plugin directory - Tied decision layer for safety Tech stack: Plain markdown Why it matters: The skills are essential for anyone looking to enhance their communication and decision-making abilities in a professional context, making it a valuable resource for developers and professionals alike.

Overview: NeuroSploit is an advanced, AI-powered penetration testing framework designed to automate and augment various aspects of offensive security operations. Key features: - Autonomous penetration testing harness - Rust-based CLI-first tool - Open-source (MIT license) - Runs on local subscriptions or API keys - Picks the right specialist agents for what it finds - Runs findings through a second model for validation - Hands the report to the user without human intervention Tech stack: Rust, CLI-first, open-source (MIT), local subscriptions or API keys Why it matters: NeuroSploit automates the security assessment process, reducing the need for human intervention and improving efficiency. It leverages AI to find more and invent less, making it a valuable tool for red teams and security professionals looking to enhance their offensive capabilities.

### Overview: Project Maya is a private GPU-based engine for running GLM-5.3-Flash and DeepSeek-V4.1-Flash-1017 models. It optimizes model performance and resource usage, ensuring fast and efficient processing on any GPU, including up to 16 cards. The engine is compatible with OpenAI- and Anthropic-compatible APIs, making it ideal for chat interfaces and advanced coding agents. Project Maya is designed to be highly customizable and efficient, with features such as multiple request handling and real-time monitoring. ### Key Features: - **Fast Processing:** Up to 118 tokens/s on four RTX 4090s, 81 on one RTX PRO 6000, and 30+ on a single RTX 5090. - **Model Compatibility:** Runs GLM-5.3-Flash and DeepSeek-V4.1-Flash-1017 models on any GPU, including up to 16 cards. - **Expert Placement:** Keeps the most important tokens for experts on the GPU, and moves less critical tokens to RAM or SSD. - **API Compatibility:** Compatible with OpenAI- and Anthropic-compatible APIs, making it suitable for various chat applications. - **Multiple Requests:** Handles up to 8 requests simultaneously, ideal for agents running multiple sessions or multiple people on the same PC. - **Resource Optimization:** Measures and optimizes GPU, CPU, RAM, and SSD usage to ensure the best performance. ### Tech Stack: Project Maya is built using

Overview: This repository contains a curated list of NLP paper summaries, aimed at making NLP techniques and topics more approachable and accessible. Contributions are welcome to add blog posts, summaries, or TL;DRs, and the community can help democratize NLP research. Key features: - Blog posts and summaries of NLP papers - TL;DRs for NLP techniques and papers - Source information for summaries - Centralized and easily accessible summaries - TL;DR section for quick refresher points Tech stack: Markdown, GitHub Why it matters: This repository serves as a reliable resource for readers, providing a gateway to the field of NLP. By democratizing NLP research, it aims to make the field more accessible and approachable. Repository: dair-ai/nlp_paper_summaries GitHub stars: 1475 GitHub description: A carefully curated list of NLP paper summaries, intended to make NLP techniques and topics more approachable and accessible.

Overview: sim-use is a cross-platform CLI that allows AI agents to observe and act on iOS Simulator and Android emulator/devices. Key features: - Observe: Turn any screen into a token-efficient outline that an LLM can reason about. - Act: Tap any element by its alias, no coordinates needed. Tech stack: Swift Why it matters: sim-use provides a compact, agent-friendly screen description and an alias-cached tap shortcut, enabling AI agents to observe and act in a few hundred milliseconds per round trip, facilitating a smooth agentic mobile development loop. GitHub stars: 1475 Repository: lycorp-jp/sim-use

### Overview: git-lrc is a free, micro AI code review tool that runs on every commit in a Git repository. It automatically checks for potential issues before code is merged, helping prevent outages, breaches, and technical debt. ### Key Features: - **Prevent Outages**: Stops AI agents from breaking your code. - **Prevent Breaches**: Ensures your code is secure and compliant. - **Prevent Technical Debt**: Reduces the risk of technical debt accumulation. - **60-Second Setup**: Easy to set up and start using. - **Full Integration**: Hooks into Git commit hooks for seamless integration. ### Tech Stack: Go ### Why It Matters: AI agents silently break things, causing outages, breaches, and technical debt. git-lrc helps prevent these issues by running an AI review on every commit before it lands in the repository. This ensures your code is secure, compliant, and free from technical debt, helping you maintain a stable and secure environment.

Overview: Refinery is an open-source tool designed for data scientists to manage and assess natural language training data. It helps in scaling, assessing, and maintaining the quality of training data as if it were a software artifact. Key features: - Semi-automate labeling of training data - Identify and flag low-quality subsets in your data - Monitor and track your data in a single interface - Integrations with popular labeling tools Tech stack: Python Why it matters: Refinery provides a data-centric approach to NLP model building, ensuring that your valuable time is spent efficiently and that your training data is of high quality, thus leading to better model performance.

### Overview: This curriculum is designed to provide foundational knowledge in machine learning, focusing on language models. It aims to equip new Elicit employees with the necessary background to effectively deploy and scale ML in production. ### Key Features: - **Introduction to Machine Learning**: Covers the basics of machine learning, including neural networks and backpropagation. - **Transformers**: Provides a deep dive into large language models, including GPT and its components. - **Key Foundation Model Architectures**: Explains key architectures like Transformers, GPT, and others. - **Training and Finetuning**: Discusses techniques for training and fine-tuning models. - **Reasoning and Runtime Strategies**: Covers in-context reasoning, task decomposition, debate, tool use, and honesty in AI. - **Applications**: Explores how ML models are applied in science, forecasting, search ranking, and more. - **ML in Practice**: Discusses production deployment, benchmarks, datasets, and advanced topics like world models, planning, uncertainty, and interpretability. - **Advanced Topics**: Covers world models, causality, planning, uncertainty calibration, active learning, and reinforcement learning. - **The Big Picture**: Discusses AI scaling, AI safety, economic and social impacts, and philosophical considerations. ### Tech Stack: Python and Jupyter Notebooks. ### Why It Matters: This curriculum is crucial for anyone interested in the field of machine learning, particularly those working with language models. It provides a solid foundation in the principles and techniques used in modern ML, making it essential for anyone looking to advance in the field