59 resources
ThinkingBox is an LLM agent framework launched by Microsoft, which is used to define MCP tools, run agent interactions and evaluate their behavior. Supports offline training data generation and reinforcement learning loops. The main functions include creating scenarios, simulating user-agent interaction and result evaluation, and is suitable for model training and evaluation.
Miles is an enterprise-oriented reinforcement learning framework designed for post-training of LLM and VLM. It supports efficient fine-tuning and optimization of multi-modal models and provides enterprise-level solutions designed to improve the performance and productivity of large-scale language and visual models.
WhisperLive is a real-time speech-to-text tool based on OpenAI Whisper, which supports microphone live broadcast and audio file transcription. It provides command line interface, Docker deployment and browser extensions. It has advanced functions such as word-level timestamps and speaker separation, and is suitable for multi-platform server-side deployment.
Maskit is a local privacy desensitization gateway specially designed for AI programming assistants. It automatically performs structured coding on sensitive information (such as keys, intranet IP, and PII) before sending a request, and streams it back when replying. Supports tools such as Cursor to ensure 100% local processing of data with zero telemetry, ensuring development privacy and security.
This warehouse provides configurations, benchmarks, and ARM64 runtime images for reproducible large language model inference on a four-node NVIDIA DGX Spark cluster through a switchless direct connection. It is designed to demonstrate high-performance deployment solutions under specific hardware topologies, including detailed performance data and verification results.
UMR is a unified action redirection code library for humanoid robots. It maps motion data from humans or other sources to the robot by learning point cloud correspondences. It mainly includes two stages: point cloud correspondence learning and correspondence-based motion optimization, supporting efficient conversion and adaptation of multiple motion source data such as bones and surfaces.
PivotHub is a Python-based CTF multi-layer intranet penetration aid. Using FastAPI backend and Vue 3 frontend, it provides intranet topology visualization, command library management and shooting range drill functions through the Web interface. Designed specifically for authorized environments, it supports localized deployment and SQLite data storage, and is suitable for teaching demonstrations and competition scenarios.
Misaki is a G2P (text-to-phone) engine designed specifically for the Kokoro model, supporting English and Japanese processing. It provides a Python library interface that can generate phoneme sequences and token mappings, and supports rollback processing through espeak, which is suitable for text preprocessing tasks before speech synthesis.
Convert PDFs, books and other materials into traceable and testable interactive learning web pages. Supports offline operations, emphasizes the separation of facts and model content, and is suitable for AI Agent environments with file access and Python execution capabilities.
CAD-MCP is a CAD control service based on Python that drives software such as AutoCAD, GstarCAD and ZWCAD for drawing through natural language instructions. It supports line, graphic drawing and layer management to achieve CAD automation without manual operations. It is mainly targeted at Windows platform users.
Provides an AI skills package for building Three.js browser games, integrating game development, UI, asset generation and automated testing. The director mode automatically coordinates various specialized skills, supports AAA-level image quality and complete workflow, and eliminates the need for manual selection of tools. It is suitable for quickly developing playable, high-quality web games.
SmartPhotoCrafter is an end-to-end automated photographic image editing framework. It automatically generates image quality analysis and performs repair (denoising, enhancement) or retouching (tonal adjustment) through closed-loop reasoning without user instructions. The tool supports high-fidelity generation and is designed to enable fully automated photo optimization and beautification.
Vibe Squad is a Python-based multi-model AI orchestration tool that defines behaviors through Markdown. It uses a coordinator to distribute tasks to 71 professional roles, execute them in parallel in isolated git workspaces, and supports independent reviews across model families (such as Claude, Codex, etc.), aiming to provide an efficient and safe automated workflow for structured coding.
TempAct is a reinforcement learning framework for autoregressive video generation, aiming to improve the rationality of action timing in videos. It jointly optimizes the LLM planner and diffusion executor to ensure that the generated video strictly follows the time sequence in the prompt words (such as pouring the sauce first and then adding the cheese), solving the problem of timing confusion in the traditional method, and is suitable for video generation scenarios that require high-precision event progress control.
ffmpeg-skill is a FFmpeg skill package for AI encoding agents (such as Claude Code, Cursor). It provides 40 tools to support media editing tasks such as video editing, audio processing, subtitle adding, and format conversion. No cloud or API Key required, purely local operation, suitable for Python 3.9+ environment.