Projects

### Overview: Instagram marketing skills for Claude Code and Codex provide captions, carousels, hooks, hashtag strategy, and a weekly plan for Claude Code and Codex. The skills are MIT-licensed and compatible with Claude Code and Codex. They draft, humanize, and publish content using Publora, and the skills are designed to optimize Claude Code and Codex's profiles and Instagram posts. ### Key features: - **Caption Writing**: Write captions in Claude Code and Codex's voice. - **Carousels**: Create Instagram carousels. - **Hooks**: Plan hooks for Claude Code and Codex's posts. - **Hashtag Strategy**: Create and optimize hashtag strategies. - **Weekly Plan**: Generate a full week of content for Claude Code and Codex. ### Tech stack: Python ### Why it matters: These skills are designed to help Claude Code and Codex create engaging and optimized Instagram content. They provide a comprehensive set of tools to draft captions, plan carousels, and optimize hashtags, ensuring that their profiles and posts are both humanized and effective. The skills are compatible with Claude Code and Codex, making them ideal for Claude Code and Codex's Instagram marketing needs.

### Overview: LiteFlowNet is a lightweight, fast, and accurate optical flow CNN designed for efficient optical flow estimation. The paper was presented at CVPR 2018 and achieved a remarkable accuracy of 3.27% on KITTI 2012 and 9.38% on KITTI 2015. ### Key Features: - **Pyramidal Features**: Captures hierarchical information for better flow estimation. - **Cascaded Flow Inference**: Combines cost volume and sub-pixel refinement for accurate flow estimation. - **Feature Warping (f-warp) Layer**: Enhances flow accuracy by warping features. - **Flow Regularization by Feature-Driven Local Convolution (f-lconv) Layer**: Improves flow accuracy and reduces computational complexity. ### Tech Stack: - **Programming Language**: C++ ### Why It Matters: LiteFlowNet significantly outperforms state-of-the-art methods like PWC-Net on various benchmarks, including KITTI 2012 and 2015, with a model size that is nearly half of PWC-Net. This makes it a valuable tool for real-time applications and research projects requiring high accuracy and efficiency in optical flow estimation.

### Overview: Cambium is a governance standard and reference toolset designed for knowledge repositories maintained with language models (LLMs). It helps operators answer essential questions about repository rules, work requirements, evidence needed, task resumption, and decision-making authority. ### Key Features: - **Profile Management**: Provides a candidate TOML profile for easy creation and safe editing. - **Task Resumption**: Enables resuming interrupted tasks with guaranteed outcomes. - **Persistent State**: Supports resumable work with coverage, required queue, and progress tracking. - **Determined Transitions**: Controls task and batch transitions, ensuring controlled writes. - **Recovery Mechanism**: Handles interruption recovery with receipts and terminal proof bindings. - **Validation Tools**: Validates global map, capability matrix, and gap register. - **Agent Interface**: Generates a consistent agent interface through subprocess consumption. - **Task Runner**: A typed task runtime runner that advances deterministic tools to the next agent. ### Tech Stack: Python ### Why It Matters: Cambium offers a structured approach to managing knowledge repositories, ensuring governance and decision-making clarity. Its tools and features facilitate effective workflow management, resumable tasks, and robust validation, making it a valuable tool for operators and maintainers of LLM-maintained knowledge corpora.

### Overview: Vibe Workflow is a free, open-source node-based AI workflow builder designed for artists and developers, offering a self-hostable alternative to Weavy AI, Krea Nodes, Freepik Spaces, and FloraFauna AI. It provides a visual, modular pipeline editor for AI-generated images and videos, eliminating vendor lock-in. ### Key Features: - **Self-Hostable**: Runs on your own server, eliminating the need for a paid subscription. - **Node-Based Editor**: A visual, modular pipeline editor for AI-generated images and videos. - **Custom AI Models**: Allows users to create and use their own AI models. - **No Subscription**: Free and open-source, with no vendor lock-in. ### Tech Stack: JavaScript ### Why It Matters: Vibe Workflow offers a powerful, open-source alternative to expensive AI creative tools, providing a self-hosted, node-based editor for artists and developers without the need for a subscription.

Overview: LARS is an application designed to run Large Language Models (LLMs) locally on your device, allowing you to upload your own documents and engage in conversations with the LLM, incorporating detailed citations to enhance accuracy and reduce AI-generated inaccuracies. Key features: - Advanced Citations: Appends detailed citations to LLM-generated responses, including document names, page numbers, text highlighting, image extraction, and a document reader. - Vast File-Formats Support: Supports PDFs, Word files (doc, docx, odt, rtf, txt), Excel files (xls, xlsx, ods, csv), PowerPoint presentations (ppt, pptx, odp), image files (bmp, gif, jpg, png, svg, tiff), Rich Text Format (RTF), HTML files. - Conversion Memory: Allows users to ask follow-up questions and resume prior conversations. - Full Chat-History: Enables users to go back and resume prior conversations. - Users can enable or disable RAG at any time via Settings. - Users can change the system prompt at any time via Settings. - Drag-and-drop for changing LLMs via Settings. - Pure llama.cpp backend - No frameworks, no Python-bindings, no abstractions - just pure llama.cpp! - Upgrade to newer versions of llama.cpp independent of LARS. - GPU-accelerated inferencing supported. - Tweak advanced LLM settings - Change LLM temperature, top-k, top-p, min-p, n-keep, set the number of model layers to be offloaded to the GPU

### Overview: This repository contains the official code for the CVPR 2022 paper "Neural RGB-D Surface Reconstruction," which introduces a novel approach for 3D reconstruction combining implicit surface representations with neural radiance fields. ### Key features: - **Code Implementation**: The repository provides the complete implementation of the paper's neural RGB-D surface reconstruction method. - **Data Format**: The dataset is structured in a specific format for easy access and processing. - **Optimization**: Users can run optimization using a provided script, which is essential for refining the model. ### Tech stack: Python ### Why it matters: This project is significant for researchers and developers interested in 3D reconstruction and surface modeling, offering a cutting-edge solution that combines traditional methods with modern neural techniques.

### Overview: Vendo is an embedded agent designed for B2B SaaS teams. It allows customers to build their own features and micro-apps directly on top of your product, rendering the UI in a brand-native sandbox. This approach ensures your source code remains untouched, making it perfect for teams looking to add bespoke features without altering their existing codebase. ### Key Features: - **Build Views:** Customers can create live views directly within the app, using the host's own components and API. - **Remix the UI:** Users can modify and apply changes to the UI components, making it easy to customize the experience. - **Automate Across Tools:** Customers can set up automation across different tools, with each action requiring approval. ### Tech Stack: TypeScript ### Why it Matters: Vendo is an open-source customization layer that empowers B2B SaaS teams to add unique features and micro-apps directly on top of their product, ensuring their source code remains untouched. This approach allows for a seamless integration and customization, making it ideal for teams looking to enhance their products without altering their existing codebase.

### Overview: This repository is dedicated to creating a state-of-the-art fully autonomous navigation and obstacle avoidance system for multi-rotor vehicles, utilizing advanced AI, computer vision, lidar, and GPS technologies. ### Key Features: - **End-to-End AI Model**: A novel AI model that directly outputs control commands from sensor inputs. - **Autonomous Navigation**: Fully autonomous navigation and obstacle avoidance system. - **Object Detection**: AI-based object detection model for tracking moving targets. - **Real-Time Flight**: Demonstrates autonomous point-to-point flight and person following capabilities. ### Tech Stack: Python ### Why it Matters: This project aims to revolutionize the field of autonomous drone technology, providing a robust and efficient solution for real-world applications. ### Repository: sieuwe1/Autonomous-Ai-drone-scripts ### Primary Language: Python ### GitHub Stars: 633 ### GitHub Description: State of the art autonomous navigation scripts using AI, Computer Vision, Lidar and GPS to control an Arducopter based quadcopter.

Overview: Ditto is a comprehensive framework designed to address the challenge of instruction-based video editing by generating high-quality video editing data. This dataset, known as Ditto-1M, is used to train the model Editto, which demonstrates superior instruction-following ability and sets a new state-of-the-art in the field. Key features: - Scalable data generation pipeline fusing image and video editing techniques. - Efficient model architecture with temporal enhancer to reduce computational overhead. - Intelligent agent for crafting diverse instructions and filtering outputs for quality control. - High-fidelity video editing dataset of one million examples. Tech stack: Python Why it matters: By providing a scalable and high-quality dataset, Ditto addresses the scarcity problem in instruction-based video editing, enabling researchers and practitioners to train advanced models like Editto, which outperform existing methods in terms of instruction-following abilities. Repository: EzioBy/Ditto Primary language: Python GitHub stars: 632 GitHub description: [CVPR'26 Highlight] Ditto: Scaling Instruction-Based Video Editing with a High-Quality Synthetic Dataset

### Overview: GMGN Agent Skills is a powerful tool designed for AI agents to query and trade multi-chain Meme tokens at high speed in real time. It provides access to trending token rankings, token fundamentals, social media signals, live trading activity, new tokens in Trenches, top holders, top traders, smart money positions, KOL holdings, insider wallets, bundled wallet exposure, and more. The tool supports market orders, limit orders, advanced take-profit/stop-loss strategy orders, and wallet management, including real-time holdings, recent P&L, and transaction history. ### Key Features: - **Real-time Data**: Data across SOL / BSC / Base / ETH is live on every query. Supports multi-parameter customization, no snapshot cache. - **Order Types**: Market Orders, Limit Orders, Take-Profit/Stop-Loss Orders, Trailing Take-Profit/Trailing Stop-Loss Orders, Multi-Wallet Batch Trading. - **Token Analytics**: Trending tokens, Token info, Token security, Token analytics, K-line (OHLCV), Wallet holdings, Monitoring & tracking. ### Tech Stack: - Python ### Why It Matters: Built for AI agents to query and trade multi-chain Meme tokens at high speed in real time. gmgn-skills gives AI agents direct access to GMGN's trending tokens, Trenches new token listings, and professional on-chain data — including Smart Money, KOL, rat trader, and bundler analytics. With 500+ professional data dimensions, you can turn your AI agent into a 24

Overview: YuLan: An Open-Source Large Language Model developed by researchers at Renmin University of China, focusing on Chinese and English language support with high-quality instruction and human preference data. Key features: - Supports Chinese and English language inputs and outputs. - Enhanced helpfulness, honesty, and harmlessness through curriculum learning. - Pre-trained on high-quality English, Chinese, and multilingual data for improved language ability. - Utilizes 4k context support for longer inputs and outputs. Tech stack: Python. Why it matters: YuLan-Chat models are designed to be highly capable and efficient, supporting Chinese and English language inputs and outputs, while also enhancing the model's helpfulness, honesty, and harmlessness through a curriculum learning strategy. This makes them suitable for a wide range of applications, including chat-based interactions, where they can provide accurate and contextually relevant responses.

### Overview: BigBird is a sparse-attention based transformer designed to handle longer sequences, significantly improving performance on NLP tasks like question answering and summarization. ### Key Features: - **Sparse Attention Mechanism**: BigBird uses sparse attention to efficiently process longer sequences. - **Theoretical Capabilities**: It leverages theoretical insights to fully utilize the capabilities of a complete transformer, enhancing performance. - **Improved Performance**: Dramatically improves performance on NLP tasks such as question answering and summarization. ### Tech Stack: Python ### Why It Matters: BigBird offers a practical solution for NLP tasks by addressing the limitations of traditional transformer models, particularly in handling longer sequences, thereby achieving better results compared to other models. ### Repository: **google-research/bigbird** - **GitHub Stars**: 632 - **Primary Language**: Python - **GitHub Description**: Transformers for Longer Sequences - **GitHub Description**: Not an official Google product. ### Code: - **Core Directory**: Contains the core components of the BigBird model. - **attention.py**: Linear attention mechanism. - **encoder.py**: Main long sequence encoder stack. - **modeling.py**: BERT and seq2seq transformer models with BigBird attention. ### Colab/IPython Notebook: - Provides a quick fine-tuning demonstration for text classification. ### Create GCP Instance: - **Project Creation**: Create a project and an instance in a specified zone. - **TPU Creation**: Create a Google Cloud TPU instance. - **SSH Access**: Access the instance using SSH.

Overview: Opytimizer is a Python library that implements various meta-heuristic optimization algorithms inspired by nature. Key features: - Genetic Algorithms - Particle Swarm Optimization - Ant Colony Optimization - Simulated Annealing Tech stack: Python Why it matters: Opytimizer provides a versatile toolkit for solving complex optimization problems efficiently and effectively, making it valuable for researchers and practitioners alike. Repository: gugarosa/opytimizer Primary language: Python GitHub stars: 632

Overview: This repository contains a crash course on deep learning for natural language processing (NLP) at ABBYY, focusing on practical applications and real-world datasets. The course is designed to be accessible and includes translations and updates to improve readability and relevance. Key features: - Sentiment analysis on the IMDB movie review dataset - Introduction to Keras - Word embeddings and their applications - PyTorch introduction and implementation of word-based machine translation - Convolutional neural networks (CNNs) - Recurrent neural networks (RNNs) for text classification and sequence labelling - Surname detector using character-level convolutions and LSTM classifiers Tech stack: Jupyter Notebook Why it matters: This course is essential for developers and researchers looking to build and apply deep learning models in NLP, providing a comprehensive guide through the process from basics to advanced topics. Repository: DanAnastasyev/DeepNLP-Course Primary language: Jupyter Notebook GitHub stars: 631 GitHub description: Deep NLP Course

### Overview: Show Me The Story is a self-hosted AI novel writer that converts ideas into outlines and chapters. It uses an OpenAI-compatible model to manage settings, plan chapters, draft and revise prose, and polish the entire book. The application is small, under 5 MB, and can be downloaded and used anywhere. ### Key Features: - Multiple projects: Chinese and English writing with independently switchable UI language. - Batch outlines, long-term direction, ending intent, and planning reviews. - Characters, worldview, organizations, relationship graphs, and story knowledge. - Chapter drafting, review, paragraph editing, targeted revision, and optional auto-confirm. - Extracted facts with source references, setting-change proposals, and foreshadow tracking. - Save knowledge and revise a chapter: selected entries enter that request in full; later writing retrieves relevant entries. - Existing-text import, optional writing/polishing skills, and a conversational assistant. - Separate final proofreading, anchored reports, per-chapter proofreading undo, and continuation projects. - Streaming output, logs, cancellation, local persistence, and text/outline/report exports. ### Tech Stack: Go ### Why It Matters: Show Me The Story is a self-hosted AI novel writer that converts ideas into outlines and chapters. It uses an OpenAI-compatible model to manage settings, plan chapters, draft and revise prose, and polish the entire book. The application is small, under 5 MB, and can be downloaded and used anywhere.

### Overview: Blended Latent Diffusion is an official implementation of the "Blended Latent Diffusion" paper presented at SIGGRAPH 2023. This project aims to accelerate the task of local text-driven editing of generic images by leveraging a recent text-to-image Latent Diffusion Model (LDM) operating in a lower-dimensional latent space. ### Key features: - **Blended Latent Diffusion Model (BLDM):** An accelerated solution to the task of local text-driven editing of generic images. - **Latent Diffusion Model (LDM):** A text-to-image model that speeds up diffusion by operating in a lower-dimensional latent space. - **Blended Diffusion:** A method to convert the LDM into a local image editor. ### Tech stack: The project is implemented using Jupyter Notebook. ### Why it matters: This project offers a faster and more precise solution for local text-driven editing of generic images, addressing the slow inference time of traditional diffusion models. It provides a more efficient and accurate way to manipulate images based on text prompts, making it particularly useful in applications requiring quick and precise image modifications.

Overview: This repository contains examples and libraries for the book "Natural Language Processing in Action," focusing on building socially responsible NLP pipelines that contribute to the communities they interact with. Key features: - Examples and libraries for the book "Natural Language Processing in Action" - Contributions to the NLP community - Code snippets and projects for practical NLP application development Tech stack: Python, Jupyter Notebook, Anaconda, Git, Bash Why it matters: This repository serves as a resource for developers interested in practical applications of natural language processing, emphasizing ethical and community-driven development practices.

### Overview: TensorSharp is a powerful .NET 10 inference engine designed for local GGUF models. It supports a console application, web-based chatbot interface, iPhone App, and Ollama/OpenAI-compatible APIs, providing a managed C# CPU kernel or native CUDA, Metal, and Vulkan backends. The engine supports Windows, macOS, iOS, and Linux, with GPU support for full capabilities. ### Key Features: - **Local Models and Interfaces:** Supports CLI, browser chat, and Ollama/OpenAI-compatible APIs. - **Text and Multimodal Models:** Handles dense and MoE GGUF models, reasoning, image/audio input, and document questions. - **Embeddings and Media Generation:** Provides text/code embeddings with optional image/audio/video inputs, and generates video with stereo audio using specific models. - **Editable Video Projects:** Offers a persistent multitrack editor in WebUX Chat and TensorAgent, native generation jobs, reviewed agent proposals, undo/redo, and local FFmpeg export. Supports reference photos for new shots. - **Efficient Serving:** Continuous batching, paged, Radix prefix-shared KV cache, and speculative decoding are enabled by default. Multi-GPU placement is available for supported models. - **Agent Skills and Code Tools:** Adds file, shell, and document workflows with the `TensorSharp.AgentHost` skill. Supports 17 bundled skills, including research, design interviews, and handoff, and allows bounded sub-agent delegation with private workspaces and read-only defaults. - **Personal Knowledge Collections:** Supports uploading files or nested folders,

### Overview: HelixWorld is a real-time interactive audio-visual world model developed by Noiz AI. It allows users to create and navigate an immersive environment where visual scenes and audio elements co-evolve in real-time. This technology enables synchronized audio and visual experiences, enhancing interactivity and immersion. ### Key Features: - Real-time interactive audio-visual world model - Visual scenes and camera-grounded spatial stereo sound co-evolve under user interaction - Continuous 6-DoF camera poses for immersive navigation - Audio is not a soundtrack laid on afterwards, but a natural part of the environment ### Tech Stack: Python ### Why It Matters: HelixWorld provides a groundbreaking solution for creating interactive and immersive audio-visual environments, making it a valuable tool for developers looking to enhance user engagement and interactivity in their projects.

### Overview: CipherChat is a novel framework designed to systematically evaluate the generalization capability of safety alignment for large language models (LLMs). It leverages human feedback and safety alignments presented in natural language to assess how effectively these models can handle non-natural languages, specifically ciphers. ### Key Features: - **Framework Design**: CipherChat uses a human-unreadable cipher to bypass safety alignments, allowing the LLMs to process inputs that are less likely to be covered by existing safety mechanisms. - **Model Evaluation**: It evaluates the model's ability to handle queries and responses in a cipher format, providing insights into the model's generalization capabilities. - **Data and Configuration**: The framework accepts various configurations and data paths, including the model name, data source, encoding method, instruction type, and demonstration toxicity. ### Tech Stack: - **Programming Language**: Python - **Dependencies**: torch, numpy, etc. ### Why It Matters: - **Enhanced Safety**: By using ciphers, CipherChat helps identify potential vulnerabilities in LLMs that could be exploited for safety alignment. - **Comprehensive Evaluation**: It provides a comprehensive evaluation of LLMs' ability to handle diverse inputs, including non-natural languages, thus improving their robustness and safety. ### Repository: - **GitHub Stars**: 630 - **Repository**: RobustNLP/CipherChat - **Primary Language**: Python ### Community: - **Twitter Discussion**: AIDB, Jiao Wenxiang - **Citation**: The paper is cited in the BibTeX format.