Projects

### Overview: AI 엔지니어 기술 면접 스터디는 2,000 이상의 GitHub 라이스타를 보유하고 있으며, 다양한 AI 기술에 대한 질문을 다루고 있습니다. 이 스터디는 통계/수학, 머신러닝, 딥 러닝 등 다양한 주제에 대한 질문을 다루고 있으며, 이를 통해 AI 엔지니어로서의 기술 스킬을 향상시키는 데 도움이 될 수 있습니다. ### Key Features: - **Interview Questions**: - **통계/수학**: 고유값과 고유벡터, 샘플링, 리샘플링, 확률 모형, 누적 분포 함수, 확률 밀도 함수, 조건부 확률, 공분산, 상관계수, 신뢰 구간, p-value, R square, 중심극한정리, 모수적 방법론, 비모수적 방법론, likelihood, probability, bootstrap, 모수적 모델, 비모수적 모델, likelihood, probability, bias, 로그 함수, 베르누이 분포, 이항 분포, 카테고리 분포, 다항 분포, 가우시안 정규 분포, t 분포, 카이제곱 분포, F 분포, 베타 분포, 감마 분포, 출장 비행기 비가 내릴 확

### Overview: DeepBI is an AI-native data analysis platform that leverages large language models to explore, query, visualize, and share data from various sources. It provides users with conversational data analysis, query generation, dashboard creation, automated reports, support for multiple data sources, and cross-platform compatibility. ### Key Features: - **Conversational Data Analysis**: Users can obtain arbitrary data results and analysis results through dialogues. - **Conversational Query Generation**: Generates persistent queries and visualizations through dialogues. - **Dashboard**: Assembles persistent visualizations into dashboards. - **Automated Data Analysis Reports**: To be developed. - **Supports Multiple Data Sources**: MySQL, PostgreSQL, Doris, StarRocks, CSV/Excel, etc. - **Multi-Platform Support**: Windows-WSL, Windows, Linux, Mac. - **International Support**: Supports Chinese and English. ### Tech Stack: Python ### Why It Matters: DeepBI is designed to redefine business intelligence by leveraging AI and large language models, enabling users to gain deeper insights into their data and make data-driven decisions. Its ability to handle multiple data sources and support cross-platform compatibility makes it a valuable tool for businesses of all sizes.

Overview: Open Computer Use is a secure cloud Linux computer powered by open-source LLMs and E2B Desktop Sandbox, designed to be controlled via keyboard, mouse, and shell commands. Key features: - Uses E2B for secure Desktop Sandbox - Operates via keyboard, mouse, and shell commands - Supports 10+ LLMs, including OS-Atlas, ShowUI, and any other models integrated - Live streams the sandbox display on the client computer - User can pause and prompt the agent - Compatible with Ubuntu, designed to work with any operating system Tech stack: Python Why it matters: Open Computer Use simplifies the integration and management of open-source language models for secure and controlled computer use, enhancing the flexibility and security of AI-powered computing environments.

Overview: Kashgari is a production-level NLP Transfer learning framework built on top of tf.keras, designed for text-labeling and text-classification tasks. It includes Word2Vec, BERT, and GPT2 Language Embedding. Key features: - Built-in pre-trained models for efficient transfer learning - Easy to use and understand code - Supports multiple NLP tasks: Named Entity Recognition, Part-of-Speech Tagging, and Text Classification - Exportable to TensorFlow Serving format for easy deployment Tech stack: Python Why it matters: Kashgari simplifies the process of applying advanced NLP models to text data, making it accessible for both academic and professional use.

### Overview: This repository serves as a personal journal for intermediate to advanced topics in machine learning, focusing on advanced concepts and practical implementations. ### Key Features: - **LLM from Scratch**: A detailed guide on building an LLM from scratch. - **LLM Hallucination**: A deep dive into LLM hallucinations and their implications. - **Graph Neural Networks (GNNs)**: Comprehensive coverage from Day 17 to Day 22. - **AlphaFold 3**: A detailed exploration of AlphaFold3, a machine learning approach to protein structure prediction. - **Large Language Models (LLMs)**: From fundamentals to advanced topics, including Day 3 and beyond. - **Graph Neural Networks (GNNs)**: Comprehensive coverage from Day 17 to Day 22. - **AlphaFold 3**: A detailed exploration of AlphaFold3, a machine learning approach to protein structure prediction. - **Quantum Machine Learning**: Exploring the intersection of quantum computing and machine learning. - **Jax**: A deep dive into the Jax library for numerical computations. - **Energy-Based Models**: A comprehensive look at energy-based models in machine learning. ### Tech Stack: Jupyter Notebook ### Why it Matters: This repository serves as a comprehensive resource for intermediate to advanced learners in machine learning, providing practical insights and implementations of cutting-edge topics.

Overview: Yuanjing Wanwu Agent Platform is a comprehensive, enterprise-grade AI agent development platform designed for secure, efficient, and compliant AI solutions. It simplifies the process of AI project delivery, making it accessible to all developers. Key features: - RAG/Knowledge Base Agent - Ontology Agent - Model Management Tech stack: Go Why it matters: Wanwu provides a full-stack FDE toolchain, simplifying every business decision and empowering every Forward Deployed Engineer (FDE) in enterprise scenarios. It offers secure, efficient, and compliant AI solutions, making it an essential tool for developers looking to build AI agents and workflows.

### Overview: HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis is a project that introduces a GAN-based model designed to generate high-fidelity speech efficiently. The model aims to improve upon existing methods by focusing on the modeling of periodic patterns within speech audio, which is crucial for enhancing sample quality. The project provides an open-source implementation and pretrained models, showcasing superior performance in terms of sample quality and speed compared to autoregressive and flow-based generative models. ### Key Features: - **HiFi-GAN Model**: A GAN-based model capable of generating high-fidelity speech. - **Efficient Sampling**: Achieves high-fidelity speech synthesis while being more efficient than autoregressive and flow-based methods. - **Periodic Pattern Modeling**: Focuses on modeling periodic patterns of speech audio to enhance sample quality. - **Human Evaluation**: Subjective human evaluation indicates that HiFi-GAN demonstrates similarity to human quality while generating 22.05 kHz high-fidelity audio 167.9 times faster than real-time on a single V100 GPU. ### Tech Stack: - **Python**: The primary programming language used for the implementation and training of the model. ### Why It Matters: HiFi-GAN is significant because it addresses the limitations of existing speech synthesis methods by focusing on the efficient and high-fidelity generation of speech audio. The model's ability to generate high-fidelity speech at a much faster rate than traditional methods makes it particularly useful in applications where real-time processing is crucial, such as in speech

### Overview: This repository is a curated list of resources for startups, providing everything from books and videos to AI tools and legal advice. The timeless advice on this list still holds, with a focus on both the fundamentals and the AI-era tooling that compresses the startup loop. ### Key Features: - Books - Videos and Courses - Essays and Reading - AI for Founders - Engineering and Infrastructure - Product and Analytics - Design and Brand - Incorporation, Legal and Equity - Finance and Banking - Security and Compliance - Venture Capital and Raising Money - Accelerators - Marketing, Sales and Metrics - Marketing Tools - Task Management and Collaboration - Hiring, Remote and People Ops - Leadership and People - Coworking and Virtual Offices - Newsletters and Podcasts - Data and Market Research - Notes - Contributing ### Tech Stack: Unknown ### Why it Matters: The timeless advice on this list still holds, with a focus on both the fundamentals and the AI-era tooling that compresses the startup loop.

Overview: AXLearn is a high-performance deep learning library built on JAX and XLA, designed for large-scale model training and deployment. Key features: - Extensive support for JAX and XLA - Object-oriented configuration system for building and maintaining models - Built-in integration with Flax and Hugging Face transformers - Scalable infrastructure for training models on thousands of accelerators - Global computation paradigm for efficient parallel processing - Support for a wide range of applications including NLP, vision, and speech recognition Tech stack: Python Why it matters: AXLearn is essential for developers looking to build, train, and deploy large-scale deep learning models efficiently and effectively. Its combination of performance, scalability, and ease of use makes it a valuable tool in the modern deep learning ecosystem.

Overview: Magic Context is a memory management system designed for coding agents, providing a seamless and long-lasting memory management experience. Key features: - Capture: Compresses the history of the project into durable knowledge. - Consolidate: Consolidates the compressed knowledge into a coherent memory. - Recall: Recalls the knowledge seamlessly, without the need for pauses or relearning. Tech stack: TypeScript Why it matters: Magic Context ensures that coding agents retain a long-term memory of the project, eliminating the need for constant relearning and maintaining the integrity of the project's knowledge over time.

Overview: modAL is a modular active learning framework designed for Python3, built on scikit-learn, that allows for rapid creation of active learning workflows with high flexibility and extensibility. Key features: - Seamless integration of scikit-learn or Keras models - Customizable query strategies and uncertainty measures - Ability to replace parts with custom solutions Tech stack: Python Why it matters: modAL provides a powerful and flexible framework for active learning, enabling developers to easily integrate their own models and tailor their workflows to specific needs, making it ideal for applications requiring efficient data labeling.

### Overview: PortalJS is an AI-native framework designed to help data teams build, develop, and ship data portals. It provides a complete solution for data infrastructure and frontend development, offering agentic skills to guide the architecture and backend setup. ### Key Features: - **AI Skills**: PortalJS includes skills like Claude Code, which can assemble and configure the portal architecture. - **Backend Support**: Supports CKAN, GitHub, and Frictionless as backend options. - **Data Loading**: Automatically loads datasets from CSV/JSON files in the `datasets.json` file. - **Frontend Development**: Provides a simple, editable Next.js frontend with no lock-in. - **Open-Source**: An open-source framework with a community-driven approach. ### Tech Stack: PortalJS uses TypeScript for its core development and includes agentic skills for backend setup. ### Why It Matters: PortalJS simplifies the process of building data portals by providing a framework that guides the architecture and backend setup. It helps teams avoid over-engineering or under-engineering by offering a modern, opinionated approach to data infrastructure. The skills provided by PortalJS can assemble and configure the entire stack, from backend to frontend, ensuring that the data is well-organized and easily accessible.

### Overview: MotionShot is a high-performance Android camera app that captures and synthesizes fast-motion sequences into a single, stroboscopic photo. Inspired by Sony's Motion Shot technology, it leverages on-device computer vision, 2D stabilization, and manual exposure controls to freeze motion in real-time without relying on cloud-based AI. ### Key Features: - **Pixel-to-Pixel Handheld Image Stabilization**: Correlates background grid pixels across frames to lock static elements in place. - **Normalized Chromaticity Motion Differencing**: Evaluates color space ratios and luminance separation for motion detection. - **High Shutter Speed Freeze-Frame Control**: Captures ultra-fast movement at shutter speeds as low as `1/8000s` ($125\,\mu\text{s}$ exposure). - **Extended ISO Gain Boost**: Maintains crisp image brightness under high shutter speeds in low light. - **Fast Motion Slide & Auto-Play Flipbook Preview**: Interactive touch-drag preview and a `▶ Play` button for instant sequence verification. ### Tech Stack: Kotlin ### Why It Matters: MotionShot bridges high-speed mobile computer vision with human kinetic expression, enabling the capture and synthesis of fast-motion sequences into a single, stroboscopic composite image. This technology is crucial for applications requiring real-time motion capture, such as sports analysis, dance choreography, and other fields where precise and detailed motion analysis is essential.

Overview: WhisperJAV is a subtitle generator for Japanese Adult Videos (JAV) that uses Qwen3-ASR, local LLM, Whisper, and TEN-VAD. It is designed to be noise-robust and runs on your own machine without cloud uploads. Key features: - Scene-based segmentation - VAD clamping - Defensive decoding and post-processing Tech stack: Python Why it matters: WhisperJAV addresses specific failure points of Whisper models, such as low signal-to-noise ratio, high density of non-verbal vocalisations, and long-form recordings, to generate subtitles that are more accurate and reliable for JAV content. Repository: meizhong986/WhisperJAV Primary language: Python GitHub stars: 2362 GitHub description: ASR/STT subtitle generator. Uses Qwen3-ASR, local LLM, Whisper, TEN-VAD. Noise-robust for JAV

### Overview: UpTrain is an open-source platform designed to evaluate and improve Generative AI applications. It provides comprehensive checks for language, code, and embedding use-cases, offering root cause analysis and actionable insights to resolve issues. ### Key Features: - **Comprehensive Checks:** Evaluates language, code, and embedding use-cases. - **Root Cause Analysis:** Identifies the root cause of failures. - **Actionable Insights:** Provides clear guidance on how to resolve issues. ### Tech Stack: Python ### Why It Matters: UpTrain is crucial for developers and researchers looking to ensure the quality and reliability of their Generative AI applications. Its ability to provide detailed feedback and actionable insights makes it an essential tool in the development process.

### Overview: ELECTRA is a method for self-supervised language representation learning, designed to pre-train transformer networks using minimal compute resources. It achieves strong results even on a single GPU, and has been shown to perform exceptionally well on downstream tasks like classification, QA, and sequence tagging. ### Key Features: - **Self-Supervised Learning**: ELECTRA models are trained to distinguish "real" input tokens from "fake" tokens generated by another neural network, similar to the discriminator in Generative Adversarial Networks (GANs). - **Efficient Training**: It can achieve state-of-the-art results on large-scale datasets like SQuAD 2.0, even when trained on a single GPU. - **Versatile Applications**: Supports fine-tuning on various downstream tasks, including classification, QA, and sequence tagging. ### Tech Stack: - **Python 3** - **TensorFlow 1.15** (or TensorFlow 2.0 in the future) - **NumPy** - **scikit-learn** and **SciPy** (for evaluation metrics) ### Why It Matters: ELECTRA provides a principled approach to pre-training transformers, making it a valuable tool for researchers and practitioners looking to improve the performance of language models without the need for extensive computational resources.

Overview: pentest-ai-agents is a collection of 50 Claude Code subagents designed to enhance offensive security research capabilities. These agents specialize in various domains such as recon, web, Active Directory, cloud, mobile, wireless, social engineering, payload crafting, reverse engineering, exploit chaining, detection engineering, and forensics. Key features: - 50 Claude Code subagents for specialized security research - AI attack-surface mapping for AI recon - Code auditor for code quality analysis - Crypto analyzer for cryptographic analysis - Password auditor for password strength analysis - Database attacker for database exploitation - Network attacker for network-based attacks - Traffic analyzer for network traffic analysis - Compliance mapper for compliance checks - Risk scorer for risk assessment - Post-exploitation set including evasion specialist, persistence planner, data exfiltrator, SCADA attacker, IOT pentester, and lateral movement Tech stack: Shell Why it matters: This toolset allows security researchers to leverage Claude Code's capabilities for advanced penetration testing, ensuring a more comprehensive and effective approach to security research and vulnerability exploitation. Repository: 0xSteph/pentest-ai-agents GitHub stars: 2358

### Overview: This repository compiles a comprehensive and up-to-date collection of datasets, tools, methods, review papers, and competitions for remote sensing change detection. It includes datasets for semantic change detection (SCD), binary change detection (BCD), disaster damage assessment (DDA), building damage assessment (BDA), and remote sensing image change captioning (RSICC). Additionally, it provides tools, methods, and resources for various aspects of remote sensing change detection, including deep learning models, traditional methods, and satellite data resources for disaster response. ### Key Features: - Datasets: Comprehensive collection of optical, multi-modal, and SAR datasets for remote sensing change detection. - Tools: Various software and tools for processing and analyzing remote sensing images. - Methods: A wide range of methods including deep learning models, traditional methods, and satellite data resources. - Review Papers: A collection of review papers on remote sensing change detection. - Competitions: Participation in remote sensing change detection competitions. - Satellite Data Resources: Resources for disaster response using remote sensing data. ### Tech Stack: - Python, R, MATLAB, and other programming languages. - Various remote sensing datasets and tools. - Deep learning models, traditional methods, and satellite data resources. ### Why It Matters: This repository serves as a valuable resource for researchers, practitioners, and students interested in remote sensing change detection. It provides a wide range of datasets, tools, methods, and competitions, making it an essential resource for anyone working in the field of remote sensing change detection.

### Overview: This project utilizes machine learning techniques to model the dynamics of high-frequency limit order books, providing a framework for capturing and predicting market movements. ### Key Features: - **Modeling High-Frequency Limit Order Book Dynamics**: Capturing the intricate patterns and behaviors of high-frequency limit order books. - **Feature Extraction**: Utilizes features such as Rise Ratio and Depth Ratio to enhance the accuracy of predictions. - **Learning Model Trainer**: Employing RandomForestClassifier, ExtraTreesClassifier, AdaBoostClassifier, GradientBoostingClassifier, and SVM to train models. - **Prediction Outcome**: Predicts the next 10 seconds using the best model, providing insights into market dynamics. - **Profit & Loss Analysis**: Evaluates the performance of trading strategies, highlighting the outcomes of various models. ### Tech Stack: Jupyter Notebook ### Why It Matters: This project offers a practical solution for high-frequency traders by providing a robust framework for capturing and predicting market dynamics, thereby enabling more informed trading decisions.

Overview: Data Designer is a powerful tool for generating high-quality synthetic datasets, allowing users to create diverse text, structured data, and images using various samplers, LLMs, image models, or seed datasets. Key features: - Generate diverse text, structured data, and images - Build multimodal workflows with image, audio, and video context - Connect generation to external tools through MCP servers and capture interaction traces - Control relationships between fields with dependency-aware generation - Validate and score outputs with Python, SQL, custom validators, and LLM judges - Extend Data Designer with plugins for specialized data sources, generation capabilities, and processing steps Tech stack: Python Why it matters: Data Designer simplifies the process of generating high-quality synthetic datasets, making it easier to create diverse and meaningful data for applications and research. Repository: NVIDIA-NeMo/DataDesigner GitHub stars: 2345 GitHub description: 🎨 NeMo Data Designer: Generate high-quality synthetic data from scratch or from seed data.