Projects

### Overview: Open-AlphaFold is an open-source implementation of AlphaFold 3, a state-of-the-art method for accurate structure prediction of biomolecular interactions. This project leverages PyTorch to achieve high accuracy and efficiency in protein structure prediction. ### Key Features: - **AlphaFold 3 Implementation**: Utilizes the AlphaFold 3 algorithm from the paper "Accurate structure prediction of biomolecular interactions with AlphaFold3". - **PyTorch Framework**: The project is built using PyTorch, enabling flexibility and ease of integration with other deep learning frameworks. - **Genetic Diffusion Module**: Implements a novel diffusion module that operates on atomic coordinates, providing a unique approach to protein structure prediction. - **Pair Representation and Single Representation**: The project includes modules for generating pair and single representation of protein structures, which are then passed through the diffusion module. - **Training and Inference**: The model is trained to predict protein structures from noisy atomic coordinates, and during inference, it uses diffusion to generate a distribution of plausible structures. ### Tech Stack: - **Python**: The primary programming language used for development. - **PyTorch**: The deep learning framework used for implementing the AlphaFold 3 algorithm and the Genetic Diffusion Module. ### Why It Matters: - **Accurate Structure Prediction**: AlphaFold 3 is known for its high accuracy in predicting protein structures, making Open-AlphaFold a valuable tool for researchers and developers in the field of structural biology. - **Versatile Application**: The project's ability to handle various protein lengths and structures makes it suitable for a

### Overview: WhisperRN is a React Native binding for Whisper and NVIDIA Parakeet ASR, built on top of whisper.cpp. This project aims to provide a seamless integration for developers working with speech recognition in their applications. ### Key Features: - **Speech Recognition Integration**: Enables developers to use Whisper and Parakeet ASR models in their React Native projects. - **Cross-Platform Support**: Works across iOS and Android platforms. - **High-Performance Inference**: Utilizes whisper.cpp for efficient speech recognition. - **Pre-built Artifacts**: Provides pre-built frameworks for easy integration, reducing the need for manual setup. ### Tech Stack: - **React Native**: The primary technology stack for building the React Native bindings. - **whisper.cpp**: The core library that powers the speech recognition functionality. ### Why It Matters: - **Ease of Use**: Simplifies the integration of speech recognition technology into existing React Native projects. - **Performance**: Leverages whisper.cpp for high-performance speech recognition, making it suitable for real-time applications. - **Cross-Platform Compatibility**: Ensures compatibility across both iOS and Android platforms, making it a versatile solution for developers. ### Repository: mybigday/whisper.rn - **Stars**: 807 - **Language**: TypeScript - **Description**: React Native bindings for Whisper and NVIDIA Parakeet ASR through whisper.cpp. - **GitHub Actions**: CI/CD pipeline for continuous integration and deployment. - **License**: MIT License - **npm Package**: Available on npm with version 1.0.0

Overview: CushyStudio is an AI and generative art platform designed for anyone to create stunning images, videos, and 3D models effortlessly. Key features: - CushyApps: A collection of visual tools for various artistic tasks. - CushyKit: A toolkit for custom app development and automation. Tech stack: TypeScript Why it matters: CushyStudio simplifies the process of creating generative AI art, making it accessible to everyone, and fosters a vibrant community of artists and developers.

### Overview: This repository is a curated list of papers that survey the latest advancements in autoregressive models in vision. It aims to provide a comprehensive overview of the field, highlighting key contributions and trends in the development of autoregressive models for generating high-quality visual content. ### Key Features: - **Unified Multimodal Models**: Papers discussing the integration of various modalities (e.g., text, images, video) to enhance generative capabilities. - **Autoregressive Diffusion-Forcing Video Generation**: Research focusing on the use of autoregressive models to generate realistic video sequences. - **Applications and Challenges**: Papers exploring the practical applications of autoregressive models and the challenges they face in real-world scenarios. ### Tech Stack: The repository uses Markdown, GitHub, and LaTeX for its documentation and presentation. ### Why It Matters: This survey is crucial for researchers and practitioners in the field of computer vision, as it provides a comprehensive overview of the latest advancements and trends in autoregressive models. It helps in understanding the current state of the art and identifying promising directions for future research. ### Repository: ChaofanTao/Autoregressive-Models-in-Vision-Survey - **GitHub Stars**: 807 - **Description**: A curated list of papers about the latest advancements in autoregressive models in vision. - **Primary Language**: Unknown - **GitHub Description**: [TMLR 2025] A survey for the autoregressive models in vision. ### Citation: Please consider citing our papers if our repository is helpful to your

Overview: OCTIS is a Python package designed to optimize and compare topic models, specifically tailored for the EACL2021 demo track. It simplifies the process of training, analyzing, and evaluating topic models using Bayesian Optimization. Key features: - Optimizes topic model hyperparameters - Supports multiple topic models including LDA, NMF, and others - Offers various evaluation metrics - Provides a web dashboard for easy control and visualization Tech stack: Python Why it matters: OCTIS simplifies the complex process of topic model optimization and evaluation, making it accessible to researchers and practitioners alike. Its integration with Google Colab and documentation on PyPI make it a valuable tool for both academic and practical applications in natural language processing and information retrieval.

Overview: `agent-skills-eval` is a test runner designed specifically for the Agent Skills ecosystem, allowing developers to empirically prove whether their skills improve the model's performance. Key features: - Runs skills against the same prompts twice, with and without the skill being loaded into context. - Provides a judge model grade for both outputs, showing a side-by-side report. - Supports OpenAI, Together, Groq, Anthropic via OpenAI-compat layers, and local Llama servers. - Offers a TypeScript SDK and CLI for custom pipelines and dashboards. - Includes deterministic checks for agents that call tools. - Generates fully spec-compliant JSON and JSONL artifacts. - Produces static HTML reports that can be published anywhere. Tech stack: TypeScript Why it matters: `agent-skills-eval` is essential for developers looking to prove the effectiveness of their skills within the Agent Skills ecosystem, ensuring that their agents are truly improving the model's performance.

Overview: This project is an implementation of a collaborative neural rendering technique that generates vivid dancing videos from hand-drawn anime character sheets (ACS). The method aims to bridge the gap between traditional animation and modern digital rendering, showcasing the potential of AI in creative arts. Key features: - Collaborative Neural Rendering using Anime Character Sheets - Ultra-Dense Pose Sequences for Two Characters - Sample Inputs of Anime Character Sheets - Streamlit-powered Web UI for easy access Tech stack: Python, Jupyter Notebook, NVIDIA GPU, CUDA, CUDNN Why it matters: This project represents a significant advancement in the field of AI and creative arts, demonstrating the potential of AI to enhance traditional animation techniques and create more lifelike and dynamic visual content.

### Overview: Amazon Skills is a collection of 53 free AI agent skills designed specifically for Amazon sellers. These skills provide expert-level analysis across various Amazon-related tasks, including keyword research, listing optimization, FBA calculations, and PPC campaign management. The skills are compatible with OpenClaw, Claude Code, Cursor, Windsurf, Codex, and any Skills-compatible agents. ### Key Features: - **Keyword Research**: Mining long-tail keywords and analyzing competition, seasonality trends, and market opportunity scores across 12 Amazon marketplaces. - **Listing Optimization**: Creating keyword-optimized listings and auditing competitor ASINs for 8-dimension scoring. - **FBA Calculator**: Calculating FBA fee breakdowns, including referral fees, fulfillment fees, and storage costs. - **PPC Campaign Builder**: Building PPC campaign structures, calculating ACoS targets, and setting bid strategies. - **Sales Estimator**: Estimating monthly sales from BSR, ASIN, or keyword across any Amazon category. - **Tariff Calculator**: Importing duties, landed costs, VAT/GST calculation for international trade routes. ### Tech Stack: - Python ### Why It Matters: Amazon Skills provides a straightforward, no-binary approach to Amazon selling, making it accessible for sellers without the need for complex setup or API keys. The skills are compatible with popular AI agents, ensuring consistent performance and integration with existing workflows.

Overview: This chatbot component kit is designed to provide a beautiful and customizable AI chatbot experience using shadcn/ui, with features like advanced attachments, rich chat interface, and fully themeable components. Key features: - 📎 Advanced Attachments: Smart preview, auto-conversion, and Markdown support. - 🧠 Thinking Process: Visual block showing LLM reasoning. - 🛠️ Tool Integration: Visual tool execution states, cancel support, and smart interrupt prompts. - 🎨 Fully Themeable: Leverages shadcn/ui's theming system. - 🌓 Dark/Light Mode: Built-in theme switching support. - 🎯 Developer Experience: Works with shadcn CLI, easy component installation, and composable API design. - 🎤 Voice Input: Speech-to-text support. - 📦 Copy & Paste Components: Install only what you need. - 🔧 Highly Customizable: Modify any aspect of the components. - 📱 Responsive Design: Works across all device sizes. - 🚀 Modern Code: Built with modern web standards. Tech stack: TypeScript. Why it matters: This kit provides a comprehensive solution for building AI chatbots with shadcn/ui, offering a rich, customizable, and fully themeable user experience. Repository: Blazity/shadcn-chatbot-kit Primary language: TypeScript GitHub stars: 805 GitHub description: 🤖 Beautifully designed chatbot components based on shadcn/ui README excerpt: A comprehensive chatbot component kit built on top of and

### Overview: Wax is a local-first shared memory layer for AI agents and on-device models, providing a single file solution for storing and retrieving text, photo, and video data. It eliminates the need for a server or API, using a pure Swift implementation on Apple Silicon. ### Key Features: - **Single File Storage**: All data is stored in a single `.wax` file, ensuring consistency and ease of access. - **Searchable Memory**: Documents, FTS5 text search, CoreML vectors, and a WAL are integrated into the file. - **Memory Consistency**: The same memories are accessible by Claude, Codex, Cursor, and Hermes. - **On-Device Tools**: Apple Foundation Models can use `waxRemember`, `waxRecall`, and `waxSearch` as on-device tools. - **Native Provider**: Hermes can use the native `wax-memory` provider for handoffs and turn sync. - **No Server Needed**: Eliminates the need for a server or API, making it ideal for mobile and IoT devices. ### Tech Stack: - Swift ### Why It Matters: - **Efficiency**: No need for a server, reducing latency and resource usage. - **Scalability**: The same memories are accessible by multiple agents, ensuring consistency. - **Ease of Use**: Pure Swift implementation makes it easy to integrate with existing Swift projects. ### Repository: - **Repository**: christopherkarani/Wax - **Stars**: 804 - **Description**: Shared Single-file memory layer for all your agents

Overview: HackerAI is an AI-powered penetration testing assistant that helps developers find and fix vulnerabilities by chatting with AI models. Key features: - Integrates with OpenRouter, OpenAI, E2B, Convex, Amazon S3, WorkOS, and Trigger.dev. - Supports optional AI models like abliteration.ai and Perplexity for enhanced security requests. - Provides a comprehensive setup script and detailed documentation. Tech stack: TypeScript Why it matters: HackerAI simplifies the process of identifying and fixing security vulnerabilities, making it easier for developers to ensure the security of their applications without needing deep technical knowledge. Repository: hackerai-tech/hackerai GitHub stars: 804

Overview: `contrastors` is a contrastive learning toolkit designed for researchers and engineers to efficiently train and evaluate contrastive models. Key features: - Built on top of Flash Attention for fast and efficient training. - Supports training on multiple GPUs. - GradCache support for large batch sizes in constrained memory environments. - Huggingface support for loading common models. - Masked Language Modeling (MLM) pretraining. - Matryoshka Representation Learning for flexible embedding sizes. - CLIP and LiT style contrastive learning. - Support for loading popular ViT models. Tech stack: Python Why it matters: `contrastors` simplifies the process of contrastive learning by providing a comprehensive toolkit that includes advanced features like Flash Attention, GradCache, and support for various pretraining techniques, making it an essential tool for researchers and engineers working on contrastive models.

Overview: Juggler is a visual workbench for AI coding agents, allowing developers to control and inspect their agent's performance locally or remotely. Key features: - Persistent conversations in tree format - Tool calls open into proper views - Inspect model transactions to see system prompt, messages, and output blocks - Operate on context, fold selected history into threads, and move items between branches - Resume the session, not merely the chat log - Stay vendor-neutral, using hosted APIs, local models, or agent CLI plugins - Keep code to yourself, traffic goes direct from your machine to the provider Tech stack: JavaScript Why it matters: Juggler provides a vendor-neutral, open-source solution for developers looking to control and inspect their AI coding agent's performance, keeping the agent where the code is and maintaining context and tooling across sessions. Repository: juggler-ai/juggler GitHub stars: 804 GitHub description: The Juggler Code Agent

Overview: Bolna is an open-source framework designed for quickly building LLM-based voice-driven conversational applications. It provides an orchestration platform, hosted APIs, and a no-code UI playground for developing AI voice agents. Key features: - Orchestration platform - Hosted APIs - No-code UI playground Tech stack: Python Why it matters: Bolna simplifies the process of creating voice AI applications, making it accessible and efficient for developers. Repository: bolna-ai/bolna GitHub stars: 801 GitHub description: Conversational voice AI agents README excerpt: Bolna is the end-to-end open-source production-ready framework for quickly building LLM-based voice-driven conversational applications. It orchestrates voice conversations using various ASR, LLM, and TTS providers and models over websockets. Bolna helps you create AI voice agents that can perform tasks such as initiating phone calls, transcribing conversations, and synthesizing responses. Bolna is designed to be open-source and accessible, making it a valuable tool for developers looking to create voice-first conversational applications.

### Overview: DAAM, or Diffusion Attentive Attribution Maps, is a cross-attention-based approach designed to interpret Stable Diffusion. This project proposes DAAM as a tool for understanding the generated images produced by Stable Diffusion models. ### Key Features: - **Cross-Attention-Based Interpretation**: DAAM uses cross-attention to interpret the images generated by Stable Diffusion. - **Stable Diffusion Compatibility**: The project supports Stable Diffusion models, including Stable Diffusion XL (SDXL). - **Cross-Attention Maps**: It generates heat maps to visualize the attention patterns in the images. ### Tech Stack: - **Python**: Main programming language used. - **Jupyter Notebook**: Primary development environment. - **PyTorch**: Used for implementing the model. ### Why it Matters: DAAM provides a novel way to understand the decision-making process of Stable Diffusion models, making it easier to interpret and debug the generated images. This is particularly useful in fields where interpretability is crucial, such as in the development of generative models or in the debugging of AI-generated content. ### Repository: - **Repository**: `https://github.com/castorini/daam` - **Stars**: 803 - **Description**: This repository contains the code for interpreting Stable Diffusion using DAAM, a cross-attention-based approach. ### Getting Started: - **Install Dependencies**: Use `pip install daam` to install the library, or clone the repository for an editable version. - **Login**: Use `huggingface-cli login`

### Overview: Video-MME is the first-ever comprehensive evaluation benchmark of Multi-modal LLMs in Video Analysis, introduced at CVPR 2025. It comprises 900 videos with 254 hours of content and 2,700 human-annotated question-answer pairs, designed to assess the capabilities of MLLMs in processing video data across various domains and modalities. ### Key Features: 1. **Duration in Temporal Dimension**: Videos range from 11 seconds to 1 hour, covering short, medium, and long-term videos. 2. **Diversity in Video Types**: Spanning 6 primary visual domains (Knowledge, Film & Television, Sports Competition, Life Record, Multilingual) with 30 subfields for broad scenario generalizability. 3. **Breadth in Data Modalities**: Integrating multi-modal inputs (video frames, subtitles, and audios) to assess the all-round capabilities of MLLMs. 4. **Quality in Annotations**: All data are human-annotated to ensure high quality and reliability. ### Tech Stack: None specified ### Why it Matters: Video-MME serves as a crucial benchmark for evaluating the capabilities of Multi-modal LLMs in video analysis, providing a comprehensive evaluation framework for researchers and developers to test and improve their models.

Overview: OmniLottie is an open-source project that generates Lottie animations using parameterized Lottie tokens and pre-trained Vision-Language Models (VLMs). It includes a multimodal dataset with two million Lottie animations and a standardized evaluation protocol for multi-modal vector animation generation tasks. Key features: - OmniLottie generates Lottie animations from multi-modal instructions. - It includes a multimodal dataset with rich annotations. - It provides a standardized evaluation protocol for multi-modal vector animation generation tasks. Tech stack: Python Why it matters: OmniLottie offers a comprehensive solution for generating detailed Lottie animations from diverse input sources, making it a valuable tool for developers and creators looking to enhance their vector animation projects.

Overview: Kiwi is a fast and versatile Korean Part-of-Speech (POS) analyzer library designed for ease of use by anyone interested in Korean language processing. It is built on the Kojin (Korean Sentence Tokenizer) model and uses the Kojin Plan and the Korean Corpus as training data. Key features: - High text analysis speed - Improved disambiguation performance - Sentence splitting capabilities Tech stack: C++ Why it matters: Kiwi is a lightweight, open-source library that accelerates Korean language processing tasks, making it an essential tool for developers and researchers working with Korean language data.

Overview: DXcam is a high-performance Python library for capturing and screen recording on Windows using the Desktop Duplication API. It offers improved FPS stability and higher frame rates compared to other Python alternatives. Key features: - Supports DirectX 12 and DirectX 11 for improved performance and compatibility. - High FPS capture, up to 240+ fps on 1080p resolution. - Stable capture for full-screen Direct3D applications. - Better FPS pacing for continuous video capture. - Dual backend support for DXGI and Windows Graphics Capture. - Seamless integration for AI agents and computer vision applications. Tech stack: Python Why it matters: DXcam is designed for low-latency, high-FPS capture pipelines, making it ideal for applications requiring real-time video capture, such as live streaming, video analysis, and AI-driven monitoring. Its superior performance and stability make it a preferred choice for developers looking to enhance their screen capture and video processing workflows. Repository: ra1nty/DXcam GitHub stars: 801 GitHub description: A Python high-performance screen capture library for Windows using Desktop Duplication API - Updated 2026

### Overview: VectorDB is a lightweight Python package designed for efficient text retrieval using embeddings and vector search. It provides a simple, local solution for storing and retrieving text content, leveraging embeddings to enhance search performance and reducing memory usage. ### Key Features: - **Local Storage**: All data is stored locally, including embeddings and vector search, ensuring maximum performance and minimal resource consumption. - **Embedding-based Retrieval**: Utilizes embeddings to quickly find relevant text chunks, making VectorDB highly efficient and scalable. - **Minimal Dependencies**: VectorDB is built with Python, making it easy to integrate into existing projects without additional dependencies. - **Flexible Options**: Offers options for custom embeddings and chunking strategies, allowing for fine-tuning to specific use cases. ### Tech Stack: - Python ### Why it Matters: - **Performance**: VectorDB excels in providing fast and efficient text retrieval, especially when dealing with large datasets. - **Scalability**: Its local storage and lightweight nature make it suitable for both small and large-scale applications. - **Ease of Use**: The package is designed to be simple and intuitive, making it accessible for developers of all skill levels. ### GitHub Stars: 801 ### Repository: kagisearch/vectordb ### Primary Language: Python ### Installation: To install VectorDB, use pip: ```bash pip install vectordb2 ``` ### Usage: ```python from vectordb import Memory # Load and search text content memory = Memory() memory.save( ["apples are