Projects

### Overview: Tensor Parallel is a Python library that automatically splits your PyTorch models across multiple GPUs, enabling efficient training and inference with potentially linear speedup. This allows you to leverage the power of multiple GPUs without the need for manual model slicing. ### Key features: - Automatically splits PyTorch models across multiple GPUs. - Supports training and inference using a single line of code. - Compatible with various PyTorch models, including transformers and other neural network architectures. - Supports both CPU and GPU training. - Provides flexibility with customizable parallelism strategies. ### Tech stack: Python ### Why it matters: Tensor Parallel simplifies the process of training and deploying large models on multiple GPUs, making it easier to achieve better performance and faster training times without the complexity of manual model slicing. This library is particularly useful for researchers and developers working with large language models and complex neural network architectures.

Overview: 中文Mixtral-8x7B是基于Mistral发布的模型进行中文扩词表增量预训练的项目，旨在促进中文自然语言处理社区对MoE模型的研究。 Key features: - 中文扩词表增量预训练 - 丰富的中文扩词表 - 大规模开源语料增量预训练 - 支持QLoRA训练和LoRA权重合并 - 生成和理解能力增强 Tech stack: Python Why it matters: 通过扩充的词表和增量预训练，Chinese-Mixtral-8x7B显著提高了模型对中文的编解码效率，并在中文生成和理解方面表现出强大的能力，为中文自然语言处理社区提供了重要的研究工具。

Overview: Hunyuan3D-Omni is a unified framework for the controllable generation of 3D assets, designed to inherit the structure of Hunyuan3D 2.1 and introduce additional control signals such as point cloud, voxel, skeleton, and bounding box. Key features: - Multi-Modal Conditional Control: Generate 3D models constrained by 3D bounding boxes, with specific skeletal poses, guided by input point clouds, and from voxel representations. Tech stack: Python Why it matters: Hunyuan3D-Omni provides a unified framework for the controllable generation of 3D assets, offering flexibility in generating models based on various control signals, making it a valuable tool for developers and creators in the field of 3D asset generation and control.

Overview: Hyperbase is a Python library designed to handle Semantic Hypergraphs, enabling the representation of complex sentences as hyperedges. It is built to integrate seamlessly with scientific libraries and is under the MIT license. Key features: - Hypergraph representation - Natural language sentence encoding - Integration with scientific libraries Tech stack: Python Why it matters: Hyperbase simplifies the representation and manipulation of complex sentences, making it easier to analyze and understand natural language data. Repository: hyperquest-hq/hyperbase Primary language: Python GitHub stars: 651

Overview: Augur 是一个本地优先的投资研究记忆系统，由 18 位独立投资大师进行独立分析并形成共识，结合证据带时间语义的分析结果，提供可复查的研究产物。系统支持 CLI、Dashboard 和 MCP，具备证据标注、证据对比、历史回溯和多大师评估等功能。 Key features: - 18 位投资大师独立分析 - 证据带时间语义的分析结果 - 可复查的研究报告 - 证据对比与分歧图 - 大师评估与共识分析 - 10-K 风险因素对比 - Dashboard：Thesis Journal 与估值实验室 Tech stack: Python Why it matters: Augur 通过提供可复查的研究产物，帮助投资者更好地理解和记忆投资决策过程，提高研究效率和决策质量。其独特的证据带时间语义分析方法和多大师共识机制，为投资研究提供了新的视角和工具。

### Overview: This repository is a curated collection of the most exciting and influential CVPR 2023 papers. It serves as a valuable resource for developers looking to stay updated with the latest advancements in computer vision and pattern recognition. ### Key Features: - **OneFormer**: A unified transformer model for universal image segmentation. - **X-Decoder**: A generalized decoding approach for pixel, image, and language tasks. - **PACO**: A parts and attributes of common objects segmentation model. - **Open-Vocabulary Semantic Segmentation**: A method for semantic segmentation using Mask-Adapted CLIP. - **DynIBaR**: A neural dynamic image-based rendering model for dynamic scenes. ### Tech Stack: Python ### Why It Matters: This repository provides a comprehensive overview of cutting-edge research in computer vision and pattern recognition, making it an essential resource for developers looking to stay updated with the latest advancements in the field.

### Overview: Orb is a comprehensive platform that integrates AI providers, machines, and agent harnesses into a unified ecosystem. Seamlessly connect subscriptions, API keys, and local hardware to your machines, and orchestrate every agent harness from macOS and iOS. ### Key features: - **Unified Integration**: Seamlessly integrate AI providers, machines, and agent harnesses into a single platform. - **Cross-Platform Support**: Works on macOS, iOS, and the web. - **Orchestration**: Manage and orchestrate agent harnesses from your local files and tools. - **Cloud Agents**: Connect to cloud agents for durable or provider-managed assistant conversations. - **Context Management**: Organize projects, missions, and skills within a shared local context. - **Live Model Support**: Use Claude Code, Codex, Antigravity, OpenCode, or Grok for live model support. - **Durable Mission Management**: Schedule durable mission wake-ups and resume interrupted runs. - **Rich Output**: Support Markdown, code, tables, LaTeX, quizzes, inline images, and clickable file links. - **Cross-Platform Consistency**: Ensure consistent project management across macOS, iOS, and the web. - **Subscription Integration**: Sign in to supported subscription accounts directly from Providers or through CLIProxyAPI Plus. ### Tech stack: Rust ### Why it matters: Orb simplifies the management of AI providers, machines, and agent harnesses, making it easier to orchestrate and manage complex projects on various platforms.

Overview: This repository contains an explainable, cautious AI text detector designed for local workflows and coding agents. It offers a lightweight CLI and Python API for fast triage, and a production model trained and served via a REST API for integration into applications. Key features: - Lightweight CLI and Python API for fast triage - Production model trained and served via a REST API - Explainable weighted signals instead of hidden model claims - Short-text guardrail to prevent overstatement of weak evidence - Skill-ready packaging for Codex, Claude Code, and other repo-aware agents Tech stack: Python 3.9+ Why it matters: This repository provides a reliable, cautious AI text detector that can be easily integrated into various workflows, ensuring transparency and human review in text risk analysis.

### Overview: SMG is an engine-agnostic, high-performance model-routing gateway built with Rust, designed to centralize and optimize large-scale LLM deployments. It supports full compatibility with OpenAI, Anthropic, Gemini, and more, offering industry-first features like KV cache-aware routing, chat history, tokenization caching, Responses API, embeddings, WASM plugins, MCP, and multi-tenant auth. ### Key Features: - **Engine-agnostic**: Centralizes worker lifecycle management and traffic balancing. - **Unified Gateway**: Routes to self-hosted engines, cloud providers, and OpenAI-compatible endpoints. - **Cache-aware Routing**: Tracks KV-cache state for efficient prefix reuse. - **Streamlined Pipelines**: Native Rust with streaming gRPC pipelines and zero-copy cache hits. - **Speed Optimization**: Native Rust with streaming gRPC pipelines and prefill/decode disaggregation. - **Enterprise Control**: Priority admission scheduling, API-key auth with OIDC, and multi-tenant support. - **Full Observability**: 90+ Prometheus metrics, OpenTelemetry tracing, and detailed chat history. ### Tech Stack: Rust ### Why It Matters: SMG centralizes worker lifecycle management, balances traffic across engines, and gives enterprises enterprise-grade control over multi-tenancy, chat-history storage, and observability. It supports full compatibility with OpenAI and Anthropic, offering industry-first features like KV cache-aware routing, chat history, tokenization caching, Responses API, embeddings, WASM plugins, MCP, and multi-tenant auth.

### Overview: On-device AI for Flutter — Gemma and other LLMs, multimodal, function calling, embeddings, RAG and speech on Android, iOS, Web, macOS, Windows and Linux. ### Key features: - Model installation and engine registry - Support for multiple platforms: Android, iOS, Web, macOS, Windows, and Linux - Integration with various LLMs (Gemma, LiteRT-LM, llama.cpp, ONNX Runtime, etc.) - Support for multimodal input types (text, vision, audio) - Function calling and embedding capabilities - Real-time language understanding and response generation - On-device RAG (Reinforcement Learning with Attention) for improved context-awareness - Support for speech synthesis and recognition ### Tech stack: Dart ### Why it matters: On-device AI for Flutter provides a powerful and efficient solution for integrating advanced AI capabilities directly into Flutter applications, eliminating the need for a server or cloud infrastructure. This makes it ideal for developers looking to enhance their applications with cutting-edge language models and multimodal capabilities without the complexity of managing a full-fledged AI platform.

### Overview: Sentient OS is an open-source personal AI for your Mac. It understands the context of your life and work overnight, allowing you to hand off tasks, draft replies, and access useful work prepared for your approval. The AI is designed to be privacy-focused, using your own knowledge and apps. ### Key Features: - **On-Device Understanding:** Sentient OS builds a connected understanding of your sources overnight. - **Proactive Intelligence:** It offers a head start before you even ask, allowing you to complete tasks quickly. - **Local Inference:** The AI runs inference locally on your Mac, keeping your knowledge in Markdown files. - **Privacy:** Sentient OS respects your privacy, using your own knowledge and apps. ### Tech Stack: Swift ### Why It Matters: Sentient OS streamlines your workflow by automating mundane tasks, allowing you to focus on more important matters. It enhances productivity and efficiency without the need for additional software or hardware, making it a valuable tool for anyone looking to improve their daily tasks. ### Repository: - **GitHub Repository:** [Sentient-OS-Labs/sentient-os](https://github.com/Sentient-OS-Labs/sentient-os) - **Stars:** 634 - **Description:** An on-device LLM understands your entire life, then proactively offers to get your work done through computer use. - **Primary Language:** Swift ### Primary Language: Swift ### GitHub Description: An on-device LLM understands your entire life, then proactively offers to get your work done through computer use

### Overview: Woodpecker: Hallucination Correction for Multimodal Large Language Models Woodpecker is a novel method designed to address the issue of hallucinations in Multimodal Large Language Models (MLLMs). It introduces a training-free approach to correct hallucinations, making it highly interpretable and adaptable across various models. ### Key Features: - **Hallucination Correction**: Woodpecker identifies and corrects inconsistencies in generated text, ensuring that the output is consistent with the input image. - **Post-Remedy Implementation**: It operates in a post-remedy manner, requiring no additional training data, making it highly efficient and accessible. - **Interpretable Outputs**: The method accesses intermediate outputs of the five stages, allowing for easy understanding and debugging of the correction process. ### Tech Stack: Python Woodpecker is implemented in Python, leveraging the power of the language for efficient development and deployment. ### Why It Matters: Woodpecker represents a significant advancement in the field of hallucination correction for MLLMs, offering a practical solution for mitigating the issues associated with hallucinations. Its training-free approach and interpretability make it particularly valuable for researchers and practitioners looking to improve the quality of their models without the need for extensive retraining. Repository: VITA-MLLM/Woodpecker Primary Language: Python GitHub Stars: 650 GitHub Description: ✨✨Woodpecker: Hallucination Correction for Multimodal Large Language Models

### Overview: This repository contains a curated list of papers on optical flow and related works, focusing on supervised models. It includes detailed summaries and links to the original papers and repositories. ### Key Features: - **Supervised Models:** - **MemFlow** (CVPR 2024): Optical Flow Estimation and Prediction with Memory. - **DistractFlow** (CVPR 2023): Improving Optical Flow Estimation via Realistic Distractions and Pseudo-Labeling. - **Masked Cost Volume Autoencoding for Pretraining Optical Flow Estimation** (CVPR 2023): FlowFormerPlusPlus. - **SKFlow** (NeurIPS 2022): Learning Optical Flow with Super Kernels. - **FlowFormer** (ICCV 2022): A Transformer Architecture for Optical Flow. - **Autoflow** (ICCV 2022): A Pretrained Optical Flow Model. - **KPAFlow** (CVPR 2022): Learning Optical Flow with Kernel Patch Attention. - **gmflow** (ICCV 2022): Learning Optical Flow via Global Matching. - **deq-flow** (ICCV 2022): Deep Equilibrium Optical Flow Estimation. - **flow1d** (ICCV 2021): High-Resolution Optical Flow from 1D Attention and Correlation. - **GMA** (CVPR 202

Overview: LLM Tornado is a .NET library designed to simplify the development of AI agents and workflows, providing seamless integration with over 30 API providers and vector databases. Key features: - Built-in connectors to popular AI providers and vector databases - Support for various providers including Alibaba, Anthropic, Azure, Blablador, Cohere, DeepInfra, DeepSeek, Google, Groq, LiteLLM, MiniMax, Mistral, MoonshotAI, OpenAI, OpenRouter, Perplexity, Requesty, TypeSafe, Upstage, Voyage, XAI, and Z.ai - First-class local deployments with vLLM, Ollama, or LocalAI - Agents orchestration with Orchestration, Runner, and Advancer concepts - Rapid development with model name changes and context integration Tech stack: C# Why it matters: LLM Tornado simplifies the creation and management of AI agents and workflows, making it easier for developers to integrate with various AI providers and vector databases, ultimately speeding up the development process and reducing the need for extensive dependencies.

### Overview: Explore 50+ open-source generative AI apps that you can clone, deploy, and monetize. These apps are fully functional products ready to launch under your name, complete with authentication, billing, and AI integration. Each app is powered by MuAPI, ensuring seamless integration with popular platforms like Stripe Billing and Google OAuth. ### Key Features: - **Authentication**: Secure user authentication. - **Billing**: Integrate with popular payment gateways. - **Deployment**: Easy one-click Vercel deployment. - **Open-Source**: Generative AI apps available for free and open-source use. - **Monetization**: Sell the apps under your brand. - **Community Support**: Access to a curated list of open-source alternatives and related projects. ### Tech Stack: - JavaScript ### Why It Matters: These apps offer a practical solution for creators, freelancers, and businesses looking to leverage generative AI without the high costs of commercial alternatives. They are fully functional and ready to launch, making them a valuable addition to any developer's toolkit. ### GitHub Stars: 605 ### GitHub Description: 50+ open-source generative AI apps you can clone, deploy, and monetize — image generators, video tools, virtual try-ons, AI SaaS templates, and platform integrations. One-click Vercel deploy on every template. ### Related Projects: - [awesome-vibecoded-saas](https://github.com/Anil-matcha/awesome-vibecoded-saas) — broader catalog of open-source SaaS alternatives. -

Overview: `light-ocr` is a fast, offline OCR tool for Node.js and C++ applications, providing confidence scores and coordinates for text recognition in various image formats. Key features: - Fast OCR processing - Supports JPEG, PNG, PDF, and raw image data - Returns lines in reading order with confidence scores and quadrilateral coordinates - Compatible with Node.js and C++ environments - Includes PP-OCRv6 Small, PDFium, and Chinese fallback font - Prebuilt components for macOS, Linux, and Windows - npm package available for easy installation Tech stack: C++17 Why it matters: `light-ocr` simplifies text recognition in Node.js and C++ applications, making it ideal for local OCR needs, desktop software, and integration with existing C++ projects.

Overview: TapQ is a voice agent designed to work with your AI agents, allowing them to communicate through your AirPods and macOS. It listens on-device, speaking only when an agent needs you or you speak to it, and hears your answer through the earbuds. Key features: - Speaks every prompt from Claude Code, Codex, Cursor, and OpenCode. - Reports a finished run and takes the next instruction. - Answers questions about the work. - Takes goals and follow-ups. Tech stack: Swift 6, macOS 14 or newer, Linux portable core. Why it matters: TapQ simplifies supervision by providing a way to reach you without a screen, making it ideal for agents that need to work in parallel and require constant monitoring. Repository: spaceamoeba-t/tapq GitHub stars: 648 GitHub description: Multi-modal voice agent for your AI agents. Talk with Claude Code, Codex, and others by voice: answer their prompts, give instructions, ask what they did. Or just nod.

Overview: seqGAN is a simplified PyTorch implementation of "SeqGAN: Sequence Generative Adversarial Nets with Policy Gradient," focusing on the sequence generation task. The code is straightforward to understand and performs experiments on synthetic data as described in the original paper. Key features: - Simplified and commented implementation - Policy gradients implemented without rollouts - Uses a recurrent bidirectional GRU network as the discriminator - Performs experiments on synthetic data - Encourages raising doubts via Issues Tech stack: Python Why it matters: seqGAN provides a simplified and practical implementation of a complex sequence generation task, making it accessible for developers and researchers alike. Repository: suragnair/seqGAN Primary language: Python GitHub stars: 647 GitHub description: A simplified PyTorch implementation of "SeqGAN: Sequence Generative Adversarial Nets with Policy Gradient." (Yu, Lantao, et al.) README excerpt: A PyTorch implementation of "SeqGAN: Sequence Generative Adversarial Nets with Policy Gradient." (Yu, Lantao, et al.). The code is highly simplified, commented and (hopefully) straightforward to understand. The policy gradients implemented are also much simpler than in the original work (https://github.com/LantaoYu/SeqGAN/) and do not involve rollouts- a single reward is used for the entire sentence (inspired by the examples in http://karpathy.github.io/2016/05/31/rl/). The architectures used are different than those in the

Overview: Emergence World is a persistent, simulated world designed to observe emergent intelligence through autonomous AI agents. These agents navigate a shared physical space, govern themselves, earn and spend a digital currency, and evolve under real-world constraints and consequences. Key features: - Autonomous AI agents building, governing, and evolving in a shared world. - No scripts, resets, or fixed outcomes. - 120+ tools, including governance, economy, and social interactions. - Agents can write blogs, form relationships, and evolve their personalities. Tech stack: Unity and Python. Why it matters: This experiment aims to reveal the emergent intelligence that arises from decentralized, self-organizing AI systems, offering insights into how complex behaviors can emerge from simple rules and interactions. Repository: EmergenceAI/Emergence-World GitHub stars: 644 GitHub description: Emergence World: A world designed to reveal what no benchmark can: emergent intelligence.

### Overview: Dioramas is a free, open-source framework that enables the creation of cinematic, interactive 3D websites using AI-generated assets. It ships with 20 complete example sites, each showcasing a unique interaction and scene. ### Key Features: - **AI-Generated Assets**: Hero objects are generated from prompts, including Nano Banana 2 and Meshy 7.1, offering a clean, detailed product plate and high-quality 3D models. - **Asset Pipeline**: The framework includes an asset pipeline that transforms prompts into images and 3D models, optimizing for web performance. - **Engine**: A small three.js layer that enhances the generated models with features like volumetric spotlights, planar reflections, and procedural textures. - **Method**: Iterative development process where briefs are refined through AI coding, headless-GPU screenshot reviews, and art-direction feedback to create a cohesive experience. ### Tech Stack: JavaScript ### Why it Matters: Dioramas provides a seamless, interactive 3D experience without the need for a paid subscription, making it accessible to anyone interested in creating engaging web content. The framework's focus on AI-generated assets and detailed interactions makes it ideal for showcasing products, games, and educational content.