Projects

Overview: Mosec is a high-performance and flexible model serving framework designed for efficient online service APIs, offering dynamic batching, pipelined stages, and cloud-friendly features. Key features: - Highly performant: Built with Rust for blazing speed and efficient CPU utilization. - Ease of use: Purely Python user interface for ML framework-agnostic model serving. - Dynamic batching: Aggregates requests for batched inference and distributes results. - Pipelined stages: Handles CPU/GPU/IO mixed workloads with multiple processes. - Cloud friendly: Designed for cloud environments with warmup, graceful shutdown, and Prometheus monitoring. Tech stack: Python, Rust, PyPI, conda-forge, PyPi, and Docker. Why it matters: Mosec simplifies the process of deploying and serving machine learning models, making it accessible and efficient for developers and businesses alike. Repository: mosecorg/mosec Primary language: Python GitHub stars: 902 GitHub description: A high-performance ML model serving framework, offers dynamic batching and CPU/GPU pipelines to fully exploit your compute machine.

Overview: Memind is a self-evolving cognitive memory and context engine designed for AI agents, providing 24/7 proactive capabilities. It leverages Java-native technology to achieve state-of-the-art performance across major benchmarks. Key features: - **State-of-the-Art Performance**: Memind excels on LoCoMo, LongMemEval, and PersonaMem benchmarks. - **Java Native Engine**: Built natively in Java, ensuring optimal performance. - **Separate User and Agent Memory**: Keeps user profiles and preferences separate from agent directives and tool experience. - **Evolutionary Insight Trees**: Transforms raw memories into structured insights for deeper understanding. - **Multi-Layer Retrieval**: Recalls the right context through various memory structures and search methods. - **Versatile Integration**: Works with multiple coding agents, local harness agents, chatbots, companions, copilots, and workflow agents. Tech stack: Java 21 Why it matters: Memind offers unparalleled performance and versatility, making it a cornerstone technology for AI-driven agents in the Java ecosystem.

Overview: A self-hosted web dashboard for the Hermes AI agent stack, providing a browser-based terminal, file explorer, session overview, cron management, system metrics, and an agent status panel behind a single password gate. Key features: - Real-time streaming chat - Agent CRUD and lifecycle management - System metrics and agent status overview - File browsing and editing - MCP server control plane - Backup and restore options - System health monitoring Tech stack: Vanilla JS + Vite · Node.js · Express · WebSocket · xterm.js · Chart.js · better-sqlite3 Why it matters: This dashboard offers a centralized management tool for the Hermes AI agent stack, streamlining agent configuration, monitoring, and access control, making it easier for teams to manage their AI infrastructure securely and efficiently. Repository: xaspx/hermes-control-interface Primary language: JavaScript GitHub stars: 900 GitHub description: A self-hosted web dashboard for the Hermes AI agent stack. Provides a browser-based terminal, file explorer, session overview, cron management, system metrics, and an agent status panel — all behind a single password gate.

**Overview:** This repository tracks the evolution of Claude Code releases, including prompts, feature flags, and metadata, providing a comprehensive history for developers. **Key features:** - `cc-prompt.md` - Legacy compatibility file per tag - `cc-flags.md` - Legacy compatibility file per tag - `system-prompts/*.md` - Extracted prompt artifacts grouped by category - `meta/flags.md`, `meta/metadata.md`, `meta/cli-surface.md`, `meta/prompt-stats.md` - Derived metadata - Annotated tags (`vX.Y.Z`) and GitHub releases for historical navigation **Tech stack:** GitHub repository **Why it matters:** This repository serves as a crucial tool for developers to track and understand the evolution of Claude Code, ensuring they can adapt and build on the latest features and improvements. **Repository:** marckrenn/claude-code-changelog **GitHub stars:** 900 **GitHub description:** Tracking prompts, feature flags, and metadata of Claude Code releases. Subscribe to ↓

### Overview: PyLate is a library designed to simplify and optimize the fine-tuning, inference, and retrieval of late interaction models, particularly focusing on ColBERT. It provides flexibility for various hardware setups and enables the construction of ColBERT models from pre-trained language models. ### Key Features: - **Fine-Tuning**: Simplifies and optimizes fine-tuning processes on both single and multiple GPUs. - **Inference**: Streamlines document retrieval and supports loading a wide range of models. - **Retrieval**: Enables efficient retrieval of documents using state-of-the-art ColBERT models. ### Tech Stack: - **Python**: Built on top of Sentence Transformers. ### Why It Matters: PyLate makes it easier to fine-tune models, optimize inference times, and retrieve documents efficiently, all while leveraging the power of pre-trained language models. This library is particularly useful for researchers and developers looking to quickly create and deploy late interaction models.

### Overview: MDistiller is a comprehensive library for knowledge distillation, providing implementations of several state-of-the-art algorithms on popular datasets like CIFAR-100, ImageNet, and MS-COCO. It includes the official implementations of both CVPR 2022 and ICCV 2023 papers on knowledge distillation. ### Key Features: - Supports KD, FitNet, AT, NST, PKT, KDSVD, OFD, RKD, VID, SP, CRD, ReviewKD, and DKD. - Implements decoupled knowledge distillation (DKD) for better performance. - Compatible with PyTorch 1.9.0 and torchvision 0.10.0. - Provides detailed benchmark results on various datasets. ### Tech Stack: Python ### Why It Matters: MDistiller offers a unified platform for researchers and practitioners to evaluate and compare different knowledge distillation methods, making it an essential tool in the field of computer vision and deep learning.

### Overview: This community-driven course on Computer Vision is a comprehensive resource for developers, covering various aspects of the field through engaging content and diverse contributions. ### Key features: - 0. Welcome - 1. Fundamentals - 2. Convolutional Neural Networks - 3. Vision Transformers - 4. Multimodal Models - 5. Generative Models - 6. Basic CV Tasks - 7. Video and Video Processing - 8. 3D Vision, Scene Rendering and Reconstruction - 9. Model Optimization - 10. Synthetic Data Creation - 11. Zero Shot Computer Vision - 12. Ethics and Biases - 13. Outlook ### Tech stack: Jupyter Notebook ### Why it matters: A unique and diverse course created by a strong open-source community, showcasing the power of collaboration and open-source projects.

Overview: Axe is a lightweight CLI tool designed for running single-purpose AI agents, allowing you to define focused agents in TOML files and execute them from various sources like cron, git hooks, or the terminal. Key features: - Multi-provider support for Anthropic, OpenAI, Ollama (local models), OpenCode, and AWS Bedrock - TOML-based agent configuration for declarative and version-controlled definitions - Sub-agent delegation for agents to call other agents via LLM tool use with depth limiting and parallel execution - Persistent memory for timestamped markdown logs that carry context across runs - Memory garbage collection for LLM-assisted pattern analysis and trimming - Skill system for reusable instruction sets that can be shared across agents - Stdin piping for direct input into agents - Local agent directories auto-discovered from `<cwd>/axe/agents/` before the global config, or use `--agents-dir` to point anywhere - Dry-run mode for inspecting resolved context without calling the LLM - JSON output for structured scripting - Built-in tools for file operations, shell command execution, URL fetching, and web search - Output allowlist for restricting `url_fetch` and `web_search` to specific hostnames - Token budgets to cap cumulative token usage per agent run via `[budget]` config or `--max-tokens` flag - MCP tool support for connecting to external MCP servers via SSE or streamable-HTTP transport - Configurable retry for transient provider errors with backoff options Tech stack: Go Why it matters: Axe

Overview: LightlyStudio is a unified data platform designed for multimodal machine learning, offering a single interface to curate, annotate, and manage various types of data, including images, videos, and more. Key features: - Integrates image, video, and text data - Supports real-time annotation and model training - Provides a user-friendly interface for data curation and management - Compatible with multiple programming languages, including Python Tech stack: Python Why it matters: LightlyStudio simplifies the process of data management and annotation, making it easier for developers to build and deploy machine learning models without the need for complex infrastructure or specialized skills. Repository: lightly-ai/lightly-studio GitHub stars: 897 GitHub description: LightlyStudio - The Unified Data Platform for Multimodal ML

Overview: OhMyCaptcha is a self-hosted captcha-solving service that provides a YesCaptcha-style API, supporting 19 task types. It is designed for use with flow2api and similar systems that expect `createTask` / `getTaskResult` semantics. Key features: - Local multimodal model (Qwen3.5-2B via SGLang) for image captcha analysis - Playwright-based solutions for reCAPTCHA v2/v3, hCaptcha, Cloudflare Turnstile - Local vision model for HCaptcha, reCAPTCHA v2, FunCaptcha, AWS grid classification Tech stack: FastAPI, Playwright, OpenAI-compatible multimodal models Why it matters: This service allows developers to integrate captcha-solving capabilities into their applications without relying on external APIs, providing a more secure and efficient solution. Repository: shenhao-stu/ohmycaptcha Primary language: Python GitHub stars: 898 GitHub description: ⚡ Self-hostable YesCaptcha-compatible captcha solver built with FastAPI, Playwright, and OpenAI-compatible multimodal models.

### Overview: This repository is a curated collection of the most exciting and influential CVPR 2025 papers. It includes highlighted papers, their abstracts, codes, videos, and demos, making it a valuable resource for developers looking to stay updated with the latest advancements in computer vision and pattern recognition. ### Key Features: - **Highlighted Papers**: 13,008 papers submitted, 2,878 accepted. - **Abstracts**: Links to the papers' abstracts. - **Codes**: Links to the corresponding codes. - **Videos**: Links to the papers' videos. - **Demos**: Links to the papers' demos. - **Session Details**: Sessions and poster sessions where the papers were presented. ### Tech Stack: Python ### Why It Matters: This repository serves as a comprehensive guide for developers interested in staying updated with the latest advancements in computer vision and pattern recognition. It provides access to the most exciting and influential papers, their codes, videos, and demos, making it an invaluable resource for anyone looking to enhance their knowledge in this field.

Overview: VLN-CE is an instruction-guided navigation task designed for continuous environments using crowdsourced instructions and realistic, unconstrained agent navigation. It is implemented using the Habitat platform. Key features: - Baseline agents and training methods for the VLN-CE task. - Support for both Room-to-Room (R2R) and Room-Across-Room (RxR) datasets. - Official implementations of "Beyond the Nav-Graph" and "Waypoint Models for Instruction-guided Navigation in Continuous Environments". Tech stack: Python 3.6, Habitat-Sim 0.1.7, Habitat-Lab 0.1.7. Why it matters: VLN-CE provides a platform for researchers and developers to interact with and benchmark their work in continuous environment navigation, offering insights into instruction-guided navigation and real-world agent capabilities.

### Overview: Nuwax Agent OS is an enterprise-grade AI Agent Development and Operation Platform that provides a comprehensive solution for agent creation, knowledge base management, model proxy, memory system, and plugin ecosystem. ### Key Features: - **Agent Creation and Distribution**: Easily build and deploy private AI solutions. - **Knowledge Base Management**: Centralized and secure knowledge management. - **Model Proxy**: Efficiently manage and access AI models. - **Memory System**: High-performance memory management for AI agents. - **Plugin Ecosystem**: Extensive ecosystem for custom agent functionalities. ### Tech Stack: TypeScript ### Why it Matters: Nuwax Agent OS simplifies the process of AI agent development and operation, making it accessible and efficient for enterprises looking to integrate AI solutions into their workflows.

Overview: This project is a local voice assistant that can operate offline, utilizing Whisper for speech-to-text conversion and ChatterBox for text-to-speech synthesis, allowing for a more immersive and authentic user experience. Key features: - Voice Cloning: Allows users to clone any voice with a short audio sample. - Emotion Control: Adjusts the emotional expressiveness of responses. - Better Performance: Utilizes a 0.5B parameter model with faster inference. - Watermarked Audio: Built-in neural watermarking for authenticity. Tech stack: Python Why it matters: This project offers a unique offline voice assistant experience, providing a more engaging and lifelike interaction compared to traditional online assistants, making it suitable for those who prefer a hands-free, offline experience. Repository: vndee/local-talking-llm GitHub stars: 896

Overview: This comprehensive book on neural networks and large language models in NLP is intended for anyone interested in NLP and deep learning. It includes chapters on foundational concepts, neural networks, sequence models, and large language models, with significant new content added. Key features: - Comprehensive coverage of neural networks and large language models in NLP - Chapters drawn from previously published articles on NLP and deep learning - Chapters cover foundational concepts, neural networks, sequence models, and large language models - Includes significant new content not present in the original articles Tech stack: Python, Jupyter Notebook, Markdown, LaTeX Why it matters: This book provides a detailed and up-to-date guide to neural networks and large language models in NLP, making it an essential resource for anyone looking to deepen their understanding of these topics. Repository: NiuTrans/NLPBook Primary language: Unknown GitHub stars: 894 GitHub description: A comprehensive book on neural networks and large language models in NLP

### Overview: PIXIU is an open-source repository that introduces the first financial large language models (LLMs) and instruction tuning data. It also provides evaluation benchmarks to assess the performance of financial LLMs comprehensively. ### Key Features: - Financial LLMs and instruction tuning data. - Evaluation benchmarks for financial LLMs. - Holistic assessment of financial LLMs. ### Tech Stack: Jupyter Notebook ### Why It Matters: PIXIU aims to continuously advance the open-source development of financial artificial intelligence (AI) by providing a comprehensive resource for financial LLMs, instruction tuning data, and evaluation benchmarks.

### Overview: This repository is a curated list of awesome tools related to artificial intelligence (AI), covering a wide range of functionalities from conversational AI to design and coding tools. It serves as a valuable resource for developers and AI enthusiasts looking to enhance their productivity and creativity. ### Key Features: - **Conversational AI:** Tools for integrating AI into chatbots and conversational interfaces. - **All-in-one tools:** Solutions for deploying AI applications across various platforms. - **AI Search Engine:** Enhances search capabilities with AI-driven suggestions. - **Writing tools:** Assist in improving writing quality and efficiency. - **Video tools:** Tools for editing, creating, and sharing videos. - **Audio tools:** Assist in editing and enhancing audio content. - **Images tools:** Tools for creating, editing, and customizing images. - **Commerce & marketing tools:** Enhances e-commerce and marketing strategies. - **Design tools:** Tools for enhancing website and app designs. - **Coding tools:** Platforms for developing AI-driven applications. - **Color tools:** Tools for enhancing color management in design. - **Miscellaneous:** Tools that don't fit into the above categories but are still useful for AI enthusiasts. ### Tech Stack: Unknown ### Why It Matters: This repository provides a comprehensive collection of tools and resources that can significantly enhance productivity and creativity in various aspects of AI and related fields. Whether you're a developer looking to streamline your workflow or an AI enthusiast exploring new tools, this list offers a wealth of possibilities.

Overview: API Relay Audit is a local security audit tool designed to detect various security issues in AI API relays and LLM proxies, including prompt injection, model substitution, tool-call rewriting, SSE anomalies, error leakage, and Web3 wallet risks. Key features: - Detects hidden prompt injection, instruction override, and context truncation. - Checks for changes in pinned package-command text, error-response leakage, and SSE stream anomalies. - Provides per-step evidence and risk assessments (LOW / MEDIUM / HIGH) for inconclusive probes. Tech stack: Python Why it matters: API Relay Audit helps developers identify and mitigate potential security vulnerabilities in their AI API relays and LLM proxies, ensuring the integrity and safety of their applications and user data.

### Overview: LaunchStack is an AI-powered startup accelerator engine built with Next.js, LangChain, PostgreSQL, and pgvector. It allows users to upload, organize, and chat with documents, including predictive missing-document detection, role-based workflows, and page-level insight extraction. ### Key features: - **Document Management**: Uploads, organizes, and searches documents. - **Chat with Documents**: Integrates with Gemini and Azure OpenAI for chat functionalities. - **Predictive Missing-Document Detection**: Automatically detects and suggests missing documents for improved organization. - **Role-Based Workflows**: Enables workflows tailored to specific roles, enhancing efficiency. - **Page-Level Insight Extraction**: Extracts detailed insights from documents, enhancing understanding. ### Tech stack: TypeScript ### Why it matters: LaunchStack simplifies the process of building AI-native applications by providing a comprehensive engine that integrates various components such as LangChain, PostgreSQL, and Next.js. This engine is designed to be modular, allowing developers to easily integrate and extend its functionalities. The use of TypeScript ensures type safety and makes the codebase more maintainable and scalable. The inclusion of features like predictive missing-document detection and role-based workflows makes LaunchStack particularly useful in accelerating the development of AI-driven applications.

### Overview: Magpie is an innovative data synthesis pipeline that generates high-quality alignment data by prompting aligned language models with their pre-query templates. Unlike traditional methods, Magpie does not require prompt engineering or seed questions, making it an efficient and effective alternative for synthetic data generation. ### Key Features: - **Prompt Engineering-Free Data Generation**: Magpie uses aligned language models to generate both the user query and the LLM response, eliminating the need for prompt engineering. - **High-Quality Synthetic Data**: The pipeline produces high-quality alignment data that can be used for various applications, including training and evaluation of language models. - **Versatile Use Cases**: Suitable for a wide range of applications, including training language models, testing hypotheses, and generating realistic dialogue data. ### Tech Stack: - **Python**: The primary programming language used for implementing Magpie. ### Why It Matters: Magpie offers a unique approach to data synthesis by leveraging the power of aligned language models, providing a more efficient and effective alternative to traditional methods. Its ability to generate high-quality data without the need for prompt engineering makes it particularly valuable for researchers and practitioners in the field of natural language processing and artificial intelligence.