A training and modeling framework for SpeechLM-based text-to-speech systems, supporting pre-training, SFT, and RLHF alignment across single or multi-GPU clusters.

DreamLite is a compact 0.39B parameter unified diffusion model that enables ultra-fast, 1024x1024 image generation and editing fully on-device.

micro_sam is a napari‑based library that adapts Meta’s Segment Anything model for microscopy. It provides interactive 2‑D/3‑D segmentation and tracking with a few clicks, includes microscopy‑fine‑tuned SAM weights, and offers optional automatic prompt generation and lightweight fine‑tuning tools.

Opentu is a canvas-based AI application platform that integrates multi-model generation and tool management into a single workspace for continuous AI task execution.

A CLI, MCP server, and AI agent that automates YouTube channel management, including video uploads, SEO optimization, and audience engagement.

A framework that uses Repository Planning Graphs (RPG) as a persistent control layer for AI coding agents to prevent requirements drift and manage complex, multi-file codebase generation and editing.

An open-source sensorimotor learning system based on the principles of the neocortex and cortical columns to enable rapid, robust learning and inference.

A collection of 22 plugins for Claude Code that add specialized agents and automation tools for development, research, and hardware design.

An AI plugin for the Fish shell that converts comments to commands, provides AI autocompletions, and suggests fixes for broken shell commands.

A pocket-sized AI chatbot device built on Raspberry Pi that enables voice-to-voice interaction, image generation, and local AI acceleration.

A Swift-based SDK and demo app that brings real-time YOLO object detection and vision tasks to iOS devices using Core ML.

A Model Context Protocol (MCP) server that connects AI assistants to DuckDB and MotherDuck for SQL analytics and data engineering.

A series of discrete diffusion Large Language Models scaled up to 100B parameters, offering high-speed parallel decoding for text generation.

WorldEngine is an open‑source post‑training platform for autonomous driving that automatically discovers rare, safety‑critical scenarios from real logs, reconstructs them with photorealistic 3‑D Gaussian Splatting, and fine‑tunes driving models via RL on the generated roll‑outs. It includes two subsystems—SimEngine for closed‑loop simulation and AlgEngine for training/evaluation—showing large gains on rare‑case benchmarks and a real‑world deployment with zero disengagements.

A Python library for generating scalable synthetic data within Databricks using Spark, designed for testing, benchmarking, and demos.

A physics-informed machine learning tool for rapid airfoil aerodynamics analysis that is significantly faster and more robust than XFoil for design optimization.

A framework for building and deploying AI agents that scales from simple chatbots to fully autonomous systems, supporting both synchronous and asynchronous workflows.

An MCP server that allows AI agents to retrieve transcripts and metadata from YouTube videos, featuring pagination to handle long videos.

A pure Go port of OpenAI's tokenizer for encoding and decoding text into tokens compatible with OpenAI's models.

A multi-tenant vector and full-text search engine that uses object storage as its primary backend with a tiered RAM/NVMe cache for low-latency retrieval.

Soar is a cognitive architecture for developing systems that exhibit intelligent behavior, providing a framework for reasoning and problem-solving.

A PyTorch-based library for training and benchmarking deep learning knowledge tracing models with standardized data preprocessing and multiple model implementations.

A React-based UI kit providing components and templates for quickly building and deploying user interfaces for Pipecat AI voice applications.

MDia is a Python library that formalises LLM intermediate reasoning as reusable “dialect cards” (LSFs) and provides a deterministic eight‑stage pipeline (collect → create → evolve → profile → select → run → validate‑rules → report) for generating, evolving, profiling, routing, and auditing these protocols across heterogeneous models. It supports black‑box LLM APIs, offers several routing plans, includes a 100‑rule bank, and guarantees reproducibility by freezing decisions before test evaluation. The repo includes a toy offline demo that runs without any API keys, extensive documentation, MIT licensing, and a citation to the ICML 2026 poster and arXiv paper.

An MCP server that integrates Ghidra's reverse engineering suite with LLM agents, enabling automated decompilation, semantic code search, and binary analysis.

Lift4D reconstructs temporally consistent 4D assets from single-view in-the-wild videos by fusing per-frame 3D estimations into a deformable model.

A multi-language SDK and reference implementation for the Agent Network Protocol (ANP), enabling AI agents to discover, authenticate, and communicate securely via standardized interfaces.

A collection of ComfyUI custom nodes optimized for the Z-Image Turbo model, offering enhanced style control, low-step consistency, and compositional variety.

Qwen3‑ASR is a self‑hosted speech‑to‑text server that wraps Qwen3‑ASR models (0.6 B & 1.7 B). It auto‑selects a GPU‑vLLM backend or a CPU‑Rust backend, offers OpenAI‑ and Alibaba‑compatible HTTP/WebSocket APIs, supports speaker diarization, VAD‑based segmentation, batch inference, and can be deployed via Docker, custom GPU images, or offline tarballs. MIT‑licensed.

A bridge that enables AI applications and agents to connect to AWS-hosted MCP servers by automatically handling AWS IAM (SigV4) authentication.

LNN is a Python library for neuro‑symbolic AI that represents logical formulas as weighted neural networks, enabling interpretable, differentiable reasoning and learning even with inconsistent or incomplete knowledge.

A local-first, multi-model AI chat client that enables group conversations between different AI models with custom personas and MCP tool support.

AgentKit is an open‑source SDK that provides a catalog of 70 ready‑made “kits” (full apps, pipelines, or templates) for building, testing and deploying AI agents. Kits are defined by a simple config file, can be edited in a visual flow studio, and are shipped as Next.js apps that run server‑less. The collection covers support triage, code review, RAG chat, hiring assistants, legal/medical bots, and many other business‑focused agents, making it a turnkey platform for creating reliable, enterprise‑grade AI agents.

MineBench is a web‑based benchmark that evaluates LLMs on 3‑D spatial reasoning by having them output voxel coordinates for prompts like “a medieval castle”. It visualises the builds, lets humans compare them head‑to‑head, and ranks models with a Bradley‑Terry system. The platform includes an Arena for pairwise voting, a Sandbox for custom builds, a Gallery of community prompts, and export options (GLB, STL, .vox, .schem). It supports major LLM providers and can be run locally with Node.js, pnpm, and Docker.

GenClaw is a code-driven agentic framework for image generation that uses executable visual sketches (SVG, HTML/CSS) as a controllable canvas before final rendering.

A managed memory layer for AI agents that provides persistent storage for facts and preferences across sessions using a two-tier session and long-term memory model.

A native-feeling port of Command & Conquer: Red Alert 2 and Yuri's Revenge for iOS and iPadOS, featuring custom RTS touch controls and a rebuilt skirmish AI.

An Obsidian plugin that integrates local AI agents (Claude Code, Codex, OpenCode) and a multi-platform publishing workbench for social media and professional documents.

A high-performance serving stack for DeepSeek-V4-Flash on NVIDIA DGX Spark, featuring continuous batching, speculative decoding, and a memory governor for massive context windows.

A Python library for LLM hallucination detection that uses various uncertainty quantification techniques to assign confidence scores to model outputs.

A general-purpose framework for reinforcement fine-tuning of LLMs that decouples exploration, training, and data buffering to support diverse RL algorithms and agentic workflows.

D-OPSD is an on-policy self-distillation framework that allows step-distilled diffusion models to be continuously tuned for new styles and concepts while preserving their fast inference capabilities.

Aliyun Model Studio CLI (bailian‑cli) is a terminal tool for Alibaba Cloud’s AI platform, providing commands for multimodal generation (text, image, video, speech), asset understanding, managed‑agent orchestration, dataset validation, fine‑tuning, deployment, and account management. Install via npm or a one‑line script, authenticate with an API key or OAuth, then issue `bl …` commands or let an AI Agent translate natural‑language prompts into CLI calls.

PhysiClaw is an AI agent that uses a robotic arm and camera to physically operate a smartphone, allowing it to automate tasks in any app without needing APIs or software installations.

MIGHTY is a ROS 2 library that plans fast, smooth trajectories using Hermite splines for UAVs, multi‑UAV swarms, and ground robots. It runs in simulation (Gazebo, RViz) or on real hardware, with Docker, native Ubuntu, and Jetson install options, and is described in an IEEE RA‑L paper.

A ROS 2 framework for creating autonomous multi-aerial-robot systems with a focus on modularity and swarm capabilities.

A plugin for Rhino 8 and Grasshopper that allows users to create, simulate, and generate manufacturer-specific code for various industrial robots.

A comprehensive introductory series and manual on robot cybersecurity, providing case studies and tutorials to highlight vulnerabilities and promote security-first robot development.