Projects

### Overview: Open-Genie is an implementation of the "Genie: Generative Interactive Environments" model, introduced by Google DeepMind. This project aims to create the first generative interactive environment trained in an unsupervised manner from unlabelled Internet videos. ### Key Features: - **VideoTokenizer**: Converts input videos into discrete tokens for the `Dynamics` module. - **LatentAction**: Models the latent space for actions. - **Dynamics**: Manages the dynamics of the environment. ### Tech Stack: - Python ### Why it Matters: Open-Genie provides a framework for creating interactive environments that can generate and respond to user actions, making it a valuable tool for developing intelligent, interactive systems.

### Overview: Securade.ai HUB is a generative AI-based edge platform designed to enhance the security and safety of CCTV cameras. By leveraging natural language text and generative AI, it eliminates the need for manual data labeling, enabling the deployment of custom, accurate computer vision models directly on the edge. This approach not only reduces costs but also accelerates the deployment of security measures. ### Key Features: - **Zero-Shot Learning**: No manual data labeling required, allowing the training of models using natural language descriptions. - **Real-Time Detection**: Live monitoring of CCTV feeds for safety and security incidents. - **PPE Detection**: Automated detection of hardhat, vest, and mask compliance. - **Proximity Alerts**: Monitor safe distances between workers and machinery/vehicles. - **Exclusion Zones**: Configurable restricted area monitoring and access control. - **Analytics Dashboard**: Real-time safety metrics and violation tracking interface. - **Multi-Camera Support**: Compatible with D-link, Tapo, TP-Link, Axis, and HikVision cameras. - **Instant Notifications**: Real-time Telegram alerts for detected violations. - **Privacy Protection**: Automatic face masking to maintain worker privacy. - **Edge Processing**: Local processing without cloud dependency. ### Tech Stack: - Python ### Why It Matters: Securade.ai HUB simplifies the process of integrating generative AI into CCTV systems, making it easier and more efficient to deploy advanced security measures without the need for extensive data labeling or cloud-based solutions. This platform is particularly useful in environments where quick deployment and real-time monitoring are crucial,

Overview: Micracode is an open-source AI web app builder that allows developers to create projects by typing descriptions and chatting with AI models. It offers a temporary preview link for public static previews and supports cloud sharing only when requested. Key features: - Temporary Preview Links - Chat-based iteration and code editing - Projects saved as plain folders - Runs in-process inside the Electron main process - Supports multiple LLM providers Tech stack: Core backend uses TypeScript and pi coding agent SDK, with Hono and Event-sourced for the backend; desktop shell is Electron 33, electron-builder for builds and updates; frontend is Next.js 15 with React and Tailwind CSS; Monaco Editor and Zustand for the editor. Why it matters: Micracode provides a lightweight, no-database, no-auth alternative to popular platforms like Lovable, Bolt, Replit, and Emergent, focusing on simplicity and ease of use.

### Overview: CrazyAra is an open-source neural network chess variant engine developed by Johannes Czech, Moritz Willig, and Alena Beyer in 2018. The project was initially created as a pure Python implementation for training a neural network to play the chess variant crazyhouse via supervised learning on human data. It was later ported to C++ by Johannes Czech, and further developed into a reinforcement learning engine that learned crazyhouse and other chess variants like chess960, King of the Hill, and Three-Check. ### Key Features: - **Deep Learning Engine:** The engine uses neural networks to analyze and predict chess moves. - **Cross-Validation:** The project includes various chess variants for cross-validation and performance testing. - **Open-Source:** The code is available for anyone to use, modify, and contribute to. ### Tech Stack: - **Programming Languages:** Python, C++ - **Frameworks:** TensorFlow, PyTorch ### Why It Matters: CrazyAra represents a significant advancement in the field of artificial intelligence and chess variants, showcasing the potential of deep learning in solving complex strategic games. Its open-source nature makes it accessible to researchers and enthusiasts alike, fostering further development and innovation in the field of artificial intelligence and chess. ### Repository: QueensGambit/CrazyAra ### Primary Language: Jupyter Notebook ### GitHub Stars: 296 ### GitHub Description: A Deep Learning Chess Variant Engine written in C++ & Python :parrot:

Overview: The Voice Assistant for Visual Studio Code is a feature-rich extension that allows users to input code snippets and execute Visual Studio commands by voice, making the development process more intuitive and accessible. Key features: - Voice commands for code snippets and project-specific commands - Support for multiple languages and project-specific voice commands - Easy reloading of voice commands - Compatibility with any Visual Studio project Tech stack: TypeScript Why it matters: This extension simplifies the development process by integrating voice commands directly into the code, making it easier for developers to work on projects without the need for a keyboard or mouse. Repository: b4rtaz/voice-assistant GitHub stars: 295 GitHub description: Voice assistant for Visual Studio Code. This extension allows you to put code snippets into the code by voice. Also, you can execute any Visual Studio command by voice. The extension works with any language, because the file with voice commands is placed in the project folder (voice-assistant.json). So, you can prepare own voice commands for each project. Repository: b4rtaz/voice-assistant Primary language: TypeScript GitHub stars: 295 GitHub description: Voice assistant for Visual Studio Code. This extension allows you to put code snippets into the code by voice. Also, you can execute any Visual Studio command by voice. The extension works with any language, because the file with voice commands is placed in the project folder (voice-assistant.json). So, you can prepare own voice commands for each project. Repository: b4rtaz/voice-assistant

Overview: EntityDB is an in-browser vector database that integrates IndexedDB, Transformers.js, and WebAssembly for efficient and scalable vector storage and retrieval. Key features: - Vector storage and retrieval - Integration with IndexedDB for persistence - Use of Transformers.js for efficient vector calculations - WebAssembly for accelerated performance Tech stack: JavaScript Why it matters: EntityDB simplifies vector database management by combining the strengths of IndexedDB, Transformers.js, and WebAssembly, offering a powerful and efficient solution for developers working with vector data in the browser.

### Overview: Generative Camera Dolly: Extreme Monocular Dynamic Novel View Synthesis is a project that introduces a novel method for generating dynamic novel views using monocular cameras. The code is part of a paper presented at ECCV 2024 and is publicly available on GitHub. ### Key Features: - **Setup Instructions**: Detailed setup instructions for using the code. - **Pretrained Models**: Main Kubric-4D checkpoints with PSNR values on the test set. - **Inference (Gradio)**: Inference code for easy visualization using Gradio. - **Dataset Processing**: Tools for processing datasets. - **Training**: Training code for generating models. - **Evaluation**: Metrics, custom controls, and evaluation of custom data without ground truth. - **Dataset Generation**: Kubric-4D, ParallelDomain-4D, and data visualization tools. - **Citations**: Links to the paper and other relevant resources. ### Tech Stack: Python ### Why It Matters: This project provides a practical solution for generating dynamic novel views using monocular cameras, which can be useful in various applications such as augmented reality, robotics, and more.

Overview: AI Talks is a user-friendly ChatGPT API wrapper that provides a Streamlit web interface, making it easy to interact with ChatGPT. Key features: - Streamlit web interface for ChatGPT API - Simplified access to advanced AI technology - Demo and token counting features Tech stack: Python Why it matters: AI Talks simplifies the interaction with ChatGPT, making it accessible and user-friendly for developers and users alike. Repository: dKosarevsky/AI-Talks Primary language: Python GitHub stars: 295

### Overview: Denser Retriever is an enterprise-grade AI retriever designed to streamline AI integration into applications, ensuring cutting-edge accuracy. It combines gradient boosting (xgboost) with keyword-based searches, vector databases, and machine learning rerankers to achieve state-of-the-art performance on the MTEB Retrieval benchmark. ### Key features: - **Supports Heterogeneous Retrievers**: Keyword search, vector search, and ML model reranking. - **xgboost ML Technique**: Combines keyword search, vector search, and reranker for improved accuracy. - **State-of-the-Art Accuracy**: Demonstrated on MTEB Retrieval benchmark. - **End-to-End Applications**: Powering chatbots and semantic search with high accuracy. ### Tech stack: Python ### Why it matters: Denser Retriever is essential for developers looking to integrate AI seamlessly into their applications, ensuring high accuracy and efficiency.

### Overview: This repository contains experiments and data related to the paper "When and why vision-language models behave like bags-of-words, and what to do about it?". The paper was presented at ICLR 2023 and received an Oral (notable-top-5%) at the conference. The datasets and models used in the paper are available for easy access and reproducibility. ### Key Features: - **Datasets**: - **Visual Genome Relation (VG-R)**: Provides relations between visual concepts and their captions. - **Visual Genome Attribution (VG-A)**: Captures the attribution of visual concepts to their captions. - **COCO-Order and Flickr30k-Order**: Order retrieval datasets for image captioning tasks. - **Models**: - **BLIP**: A pre-trained vision-language model. - **CLIP**: A pre-trained vision-language model. - **Flava**: A pre-trained vision-language model. - **XVLM**: A pre-trained vision-language model. ### Tech Stack: The repository uses Python for all experiments and data processing. ### Why It Matters: This project provides a comprehensive set of tools and datasets to understand and improve the behavior of vision-language models, particularly in their ability to capture and reproduce the concepts of visual and textual data as if they were a bag of words. The results of this work can be crucial for researchers and practitioners looking to enhance the performance of vision-language models in various applications.

Overview: PocketMen with You is a project that allows developers to create high-fidelity Codex companions from two or more photos using local open-weight neural editing without requiring an OpenAI API key. Key features: - Neural Local Studio: Uses an open-weight multi-reference image editor to create a canonical master and state-specific animation key poses. - Deterministic v0.2 engine: Provides a reliable and private engine that can only warp supplied pixels. - No OpenAI API key: Achieves results that are closer to a modern hosted image editor while remaining independent from OpenAI ImageGen. Tech stack: Python Why it matters: This project offers a solution for developers who want to create high-quality Codex companions without the need for an OpenAI API key, making the process more accessible and private. Repository: six-nut/PocketMen-with-you GitHub stars: 295 GitHub description: Create high-fidelity Codex companions from 2+ photos with local open-weight neural editing and no OpenAI API key.

### Overview: ezdata is an AI-native data platform designed to unify heterogeneous data sources, integrate ETL, orchestrate tasks, build per-source knowledge bases (RAG), and provide AI analysis capabilities. It is built on Python and supports multiple connectors, including RDBMS, Elasticsearch, MongoDB, Kafka, and vector stores, among others. ### Key Features: - **Data Management**: Supports 60+ connectors for various data sources, including RDBMS, Elasticsearch, MongoDB, Kafka, and vector stores. It also includes ETL integration built on dlt, read-only guardrails, and AI-driven querying. - **Task-Level Lineage**: Provides a declarative lineage graph for impact analysis and staleness protection, with real-time updates and single-node/distributed run modes. - **Knowledge Base (RAG)**: Extracts documents from various formats, chunks them, and uses Elasticsearch vector store for contextual retrieval, incremental training, and dedicated knowledge bases per data source. - **AI**: Unified AI model management, reasoning/thinking models supported, and an AI chat agent for data discovery, schema inspection, and query execution within a sandbox, along with agent skills and AI tools. ### Tech Stack: Python ### Why It Matters: ezdata offers a comprehensive solution for AI-native data management, enabling efficient data ingestion, integration, and analysis. Its features make it particularly useful for organizations looking to leverage AI for data retrieval and analysis, ensuring data is consistent, up-to-date, and contextually relevant.

### Overview: AI Closet is an AI-native app designed to digitize your wardrobe, providing outfit inspiration and virtual try-on functionality. The app supports both iOS and Android, leveraging React Native and Expo for cross-platform development. ### Key Features: - **Digitize Your Closet:** Easily add and manage all your clothing items from the camera or gallery. - **Virtual Try-On:** Upload a personal photo to generate an approximate look, and track previous try-on sessions. - **Outfit Management:** Mix and match items on a freeform canvas, save and revisit your favorite outfits, and list items in each outfit. ### Tech Stack: - **TypeScript:** Primary language for safer code and better autocompletion. - **React Native & Expo:** Cross-platform mobile development for iOS and Android. - **AsyncStorage & expo-file-system:** Local data persistence and file handling. - **AWS Services (Planned):** API Gateway, Lambda, DynamoDB, S3 for future cloud integration. ### Why It Matters: AI Closet simplifies the process of managing your wardrobe and provides creative outfit inspiration, making it a valuable tool for fashion enthusiasts and those looking to stay organized.

### Overview: Mol-Instructions is a large-scale biomolecular instruction dataset designed for large language models. It consists of three main components: - **Molecule-oriented instructions:** 148,4K instructions across six tasks. - **Protein-oriented instructions:** 505K instructions across five distinct categories. - **Biomolecular text instructions:** Facilitates protein design based on textual directives. ### Key Features: - **Molecule-oriented instructions:** Focuses on small molecules, covering diverse chemical reactions and molecular design challenges. - **Protein-oriented instructions:** Addresses tasks related to protein structure, function, and activity prediction. - **Biomolecular text instructions:** Enables the design of proteins using textual instructions. ### Tech Stack: - Python ### Why It Matters: Mol-Instructions provides a comprehensive dataset for training and evaluating large language models, particularly in the context of biomolecular instructions and protein design.

### Overview: AdcSR is a real-world image super-resolution (SR) method that achieves state-of-the-art performance by removing and pruning redundant modules from the one-step diffusion network and applying adversarial distillation to retain generative capabilities. It achieves super-resolution in just 0.03 seconds on an A100 GPU, surpassing the performance of OSEDiff while reducing parameters by 74%. ### Key features: - **Adversarial Diffusion Compression (ADC):** Reduces redundant modules and prunes prunable modules to maintain generative capabilities. - **Real-Time Stable Diffusion-Based Image Super-Resolution:** Super-resolves a 128×128 image to 512×512 in just 0.03 seconds on an A100 GPU. - **Competitive Visual Quality:** Achieves high visual quality across multiple benchmarks despite 74% fewer parameters. ### Tech stack: PyTorch ### Why it matters: AdcSR is a significant advancement in the field of image super-resolution, offering high performance with minimal computational resources, making it ideal for real-world applications where speed and efficiency are crucial.

### Overview: This repository is a curated list of resources on cold-start recommendation (CSR), focusing on the era of large language models (LLMs). It aims to provide a comprehensive literature base for the research and industrial community of cold-start recommendation. ### Key Features: - **Data Incomplete Learning**: Robust Co-Training, Knowledge Alignment, Cold Exploration, Feature Similarity Measurement. - **Data Efficient Learning**: Meta-Learning Optimization, Meta-Task Utilization, Meta-Embedding Initialization, Sequential Meta-Learning. - **Graph Relations**: Interaction Graph Enhancement, Homophily Network Relation, Heterogeneous Graph Relation, Attributed Graph Relation, Knowledge Graph Relation. - **Domain Information**: Domain Knowledge Transfer, Heterogeneous Connection, Learning Process, Collaborative Filtering Alignment, Auxiliary Feature Alignment, Domain-Invariant Representation Learning. - **World Knowledge from LLMs**: Prompting LLMs for Recommendation, Efficient Tuning LLMs for Recommendation, LLMs for Representation Enhancement, LLMs for Relation Augmentation. ### Tech Stack: Unknown ### Why It Matters: This repository serves as a valuable resource for researchers and practitioners interested in cold-start recommendation, providing a comprehensive overview of the field and its recent advancements.

Overview: TopMost is a comprehensive toolkit for topic modeling, covering various scenarios such as basic, dynamic, hierarchical, and cross-lingual topic modeling. It includes models, evaluation metrics, and datasets, aiming to simplify the lifecycles of topic modeling. Key features: - Basic Topic Modeling - Dynamic Topic Modeling - Hierarchical Topic Modeling - Cross-Lingual Topic Modeling Tech stack: Jupyter Notebook Why it matters: TopMost provides a unified platform for topic modeling, making it easier to implement and evaluate different models and scenarios. Repository: bobxwu/TopMost Primary language: Jupyter Notebook GitHub stars: 294 GitHub description: [ACL 2024 Demo] A Topic Modeling System Toolkit README excerpt: .. image:: https://github.com/BobXWu/TopMost/raw/main/docs/source/_static/architecture.svg :width: 390 :align: center +------------------------------+---------------+--------------------------------------------+-----------------+ | Scenario | Model | Evaluation Metric | Datasets | +==============================+===============+============================================+=================+ | | | LDA_ | | | | | | NMF_ | | | 20NG | | | | ProdLDA_ | | TC | | IMDB | | | | DecTM_ | | TD | | NeurIPS | | | Basic Topic Modeling | | ETM_

### Overview: Transfer NLP is a comprehensive library designed for reproducible experimentation and transfer learning in Natural Language Processing (NLP). Built on top of PyTorch, it simplifies the process of setting up and running experiments, ensuring consistent results and facilitating easy ablation studies. ### Key Features: 1. **Reproducible Experiment Management**: The library provides an experiment builder that enforces the use of configuration files for experiments, ensuring consistent and reproducible results. 2. **PyTorch Support**: Built on PyTorch, Transfer NLP supports various deep learning models and optimizes them for transfer learning. 3. **Tensorboard Monitoring**: All DL-based examples on the provided Colab notebooks include in-cell Tensorboard training monitoring, enhancing visualization and debugging. 4. **Customizable Object Registration**: The library allows you to register any class, method, or function, making it highly flexible and customizable. 5. **Configuration Files**: Supports multiple formats like Python Dictionary, JSON, YAML, and TOML for defining experiments, ensuring compatibility and ease of use. ### Tech Stack: - **Python**: The primary language used for development and experimentation. - **PyTorch**: The deep learning framework integrated into Transfer NLP for efficient model training and transfer learning. ### Why It Matters: - **Consistency and Reproducibility**: Ensures that experiments are run consistently across different environments, leading to reliable results. - **Ease of Use**: Simplifies the setup and management of experiments, making it easier for researchers and developers to focus on the core problem at hand. - **Customization**:

Overview: Guaardvark is a self-hosted AI studio that allows users to interact with their files, agents, and media through a local GPU. It eliminates the need for cloud API keys and per-token fees, ensuring complete privacy and control over the AI experience. Key features: - Chat with your files using RAG (Retrieval-Augmented Generation) - Hand tasks to agents that use a real desktop and browser - Run parallel coding agents - Generate video, images, music, and voice locally Tech stack: Python Why it matters: Guaardvark provides a secure, self-hosted alternative to cloud-based AI services, allowing users to control their AI interactions and avoid any telemetry or data sharing.

### Overview: 本项目是一个轻量级的本地知识库 RAG 问答系统，使用 Streamlit、LangChain、Chroma 和 Chroma 进行构建。知识库支持上传 `.txt` 文件，并通过 RAG 检索增强问答系统进行提问和回答。系统具备知识库切分、MD5 去重、流式输出和会话历史管理等功能。 ### Key features: 1. 知识库上传：通过 Streamlit 页面上传 `.txt` 文件，自动切分并写入 Chroma 向量库。 2. RAG 问答：采用 Streamlit 聊天界面，由模型基于检索到的知识库内容进行回答。 3. 会话管理：支持消息历史留存和基于历史的连续追问功能。 ### Tech stack: - Python 3.10 - Streamlit - LangChain - Chroma - FAISS (Chroma 向量库的替代方案) ### Why it matters: 适合作为 RAG 入门实践，依赖轻量级、无需 GPU，替换为自己的业务文本即可使用。适用于电商客服场景，提供知识库上传与检索增强问答的功能。 ### Repository: - **lhh737/KnowledgeBase-RAG-LLM-System** - **GitHub stars**: 292 - **GitHub description**: 轻量级本地知识库 RAG 问答系统 · Streamlit + LangChain + Chroma，支持知识库上传与检索增强问答 ### Primary language