Projects

### Overview: This repository tracks the latest advancements in Natural Language Processing (NLP), knowledge graphs, large language models (LLMs), and agent memory. It includes research and application summaries, resources for knowledge graph construction, and discussions on the practical applications of these technologies in AI agents. ### Key features: - Research and application summaries on NLP and knowledge graphs - Resources for constructing knowledge graphs and related algorithms - Discussions on the impact of NLP and knowledge graphs on AI agents ### Tech stack: JavaScript ### Why it matters: This repository provides valuable insights into the intersection of NLP, knowledge graphs, and AI agents, making it a valuable resource for developers and researchers in the field.

### Overview: This repository contains a curated list of AI Engineering interview questions and answers, organized by company. It covers a wide range of topics such as LLM internals, inference, safety, multimodal AI, and more. ### Key Features: - **Company-Wise Questions**: Questions and answers are categorized by company, making it easy to find relevant information. - **Comprehensive Coverage**: Questions cover various aspects of AI engineering, including LLM internals, inference, safety, multimodal AI, and more. - **Answers Linked**: Where available, answers are linked to provide additional context or explanations. ### Tech Stack: Markdown ### Why It Matters: This cheat sheet is invaluable for candidates preparing for AI Engineering interviews at top companies. It helps in understanding the technical landscape and staying updated with the latest trends in AI. ### Repository: pallavi-shekhar/ai-engineering-interview-questions-company-wise ### GitHub Stars: 1760 ### GitHub Description: Your Cheat Sheet For AI Engineering Interviews at Top AI Companies - Questions and Answers.

### Overview: React Native ExecuTorch is a high-performance, privacy-first on-device AI inference library for React Native, powered by Meta's ExecuTorch runtime. It enables running machine learning models directly on the user's phone with zero network calls, ensuring full offline capability and guaranteed privacy. ### Key Features: - **Native Hardware Acceleration**: Optimized execution delegates across backends: **XNNPACK** (CPU), **ARM NEON** (GPU), and **OpenCL** (GPU). - **Privacy-Preserving Inference**: Ensures that no data leaves the device, maintaining user privacy. - **Pre-Exported Models**: Includes a curated set of pre-exported models covering computer vision, language models, text-to-speech, transcription, and more, available in the Hugging Face collection. - **Custom Models**: Allows bringing your own `.pte` models or building custom models from scratch. - **Interactive Showcase App**: Available in the React Native ExecuTorch Gallery for exploring on-device capabilities. ### Tech Stack: TypeScript ### Why It Matters: React Native ExecuTorch simplifies the process of deploying AI models on mobile devices, providing a seamless and privacy-protected experience. It enables developers to leverage the power of machine learning without compromising user privacy, making it ideal for applications requiring real-time inference capabilities on mobile platforms.

### Overview: This project is the first-place solution for the CrowdFlower Product Search Results Relevance Competition on Kaggle, achieving a Public LB score of 0.70807 and a Private LB score of 0.72189. ### Key Features: - A median ensemble of 35 best Public LB submissions. - An XGBoost model with a Public LB score of 0.69322 and a Private LB score of 0.70768. - A comprehensive documentation and flowchart available for reference. ### Tech Stack: C++ ### Why It Matters: This project demonstrates the effectiveness of ensemble methods and XGBoost in improving model performance, showcasing the importance of robust feature engineering and model optimization in Kaggle competitions.

### Overview: PaddlePaddle/Research 是一个基于飞桨（PaddlePaddle）的前沿研究项目，涵盖了计算机视觉、自然语言处理、知识图谱和时空数据挖掘等领域的顶会论文和比赛冠军模型。项目旨在展示飞桨在深度学习领域的应用和创新成果。 ### Key features: - **计算机视觉**: 图像检索、车流统计、车辆再识别、车辆异常检测、医学图像分析、光流估计、语义分割、轻量化检测、地标检索与识别、图像分类、图像分类、小样本学习、迁移学习。 - **自然语言处理**: 中文词法分析、主动对话、语义解析、多轮对话、阅读理解、关系抽取、机器翻译、阅读理解、对话生成。 - **知识图谱**: 无具体信息。 - **时空数据挖掘**: 无具体信息。 ### Tech stack: Python ### Why it matters: 该项目展示了飞桨在深度学习领域的广泛应用和创新成果，为科研和开发提供了宝贵的资源和经验。

### Overview: AngelSlim is a model compression toolkit designed to enhance usability, comprehensiveness, and efficiency. It compresses large models without significant performance degradation, making it ideal for both research and production environments. ### Key Features: - **Model Compression**: AngelSlim reduces the size of large models while maintaining their performance. - **Comprehensive Documentation**: Detailed documentation available on Hugging Face and AngelSlim's official website. - **Accessibility**: The toolkit is designed to be easily accessible and understandable. ### Tech Stack: Python ### Why It Matters: AngelSlim is a valuable tool for researchers and developers looking to optimize model sizes without compromising performance. Its comprehensiveness and efficiency make it a go-to solution for various applications, including research and production environments. ### Repository: Tencent/AngelSlim ### GitHub Stars: 1769 ### Latest News: - [26/09/01] AngelSlim released the Hy4 preview MIX-STQ1_0 version, reducing model size by 99.3% with a 0.7% performance drop on SWE-bench Pro. - [26/07/29] AngelSpec, a training framework for speculative decoding, was open-sourced, with DFly and MTP + TTT methods delivering up to 2.40× average end-to-end speedup over AR and +15.7% throughput at high concurrency. - [26/07/20] Support for scale-only quantization-aware distillation on Megatron-Core for Qwen3-M

### Overview: Klee is a macOS-native AI chat application powered by MLX, running entirely on Apple Silicon with no cloud or API requirements. It features 100% local inference, no account or subscription, and supports native tool calling for file operations, web searches, and shell commands. ### Key features: - **100% local inference**: Your data stays on your Mac. - **No account or API key required**: Download and start chatting. - **One-click model download**: Pick a model, Klee handles the rest. - **Native tool calling**: The AI can read/write files, search the web, and run shell commands. - **Vision support**: Attach images to your messages with supported VLM models. - **Web search**: Search the web via Jina AI (free API key, configurable in sidebar). - **Streaming responses**: Tokens appear as they're generated. - **Inline thinking**: See the model's reasoning process in a collapsible card. - **Platform modules**: Extend the AI with native Swift integrations (coming soon). - **Lightweight**: ~75MB native SwiftUI app, no Electron, no Docker, no background services. - **System requirements**: macOS 15.0 (Sequoia) or later, Apple Silicon (M1 or later), 16 GB RAM or more. ### Tech stack: Swift ### Why it matters: Klee is designed for macOS users who prefer a local, privacy-focused AI experience. It leverages MLX for efficient model inference on Apple Silicon, ensuring that your conversations stay

### Overview: Reverse API Engineer is a powerful tool that converts websites into clean, typed APIs. It allows developers to easily integrate website functionality into their applications without manually copying and pasting code snippets. ### Key Features: - **Automatic API Generation**: Converts website traffic into Python, JavaScript, TypeScript, Go, Java, C#, PHP, Ruby, C, and PowerShell client libraries. - **Browser Automation**: Supports Playwright MCP, Chrome DevTools, and Vercel agent-browser CLI for capturing network traffic. - **Manual Mode**: Allows you to manually drive the browser and generate the client from captured traffic. ### Tech Stack: Python ### Why It Matters: - Simplifies API integration by automatically generating client libraries from website traffic. - Reduces manual effort required for API development. - Useful for developers who frequently need to integrate website functionality into their applications. ### Repository: - **Repository**: kalil0321/reverse-api-engineer - **Stars**: 1769 - **Description**: The agent that turns websites into APIs! Browse or let an AI agent browse, and get a clean, typed client for the endpoints the site actually uses. ### Installation: - **Install**: `uv tool install reverse-api-engineer` or `pip install reverse-api-engineer` - **Agent Mode**: Supports Playwright MCP, Chrome DevTools, and Vercel agent-browser CLI. - **Manual Mode**: Supports local Playwright browser, requires `playwright install chromium`. ### Quick Start: ```bash reverse-api-engineer > fetch all apple jobs from their

Overview: This project presents an efficient implementation of Wavenet generation using dynamic programming, significantly reducing the computational complexity from \(O(2^L)\) to \(O(L)\) by eliminating redundant convolution operations. Key features: - Efficient Wavenet generation algorithm - Reduces computational complexity - Provides an example implementation Tech stack: Python Why it matters: This implementation is crucial for speeding up Wavenet generation, making it more practical for real-time applications and large-scale data processing.

Overview: AlphaPy is a Python-based machine learning framework designed for both traders and data scientists. It utilizes libraries such as scikit-learn, pandas, and others for feature engineering and visualization. Key features: - Run machine learning models using scikit-learn, Keras, xgboost, LightGBM, and CatBoost. - Generate blended or stacked ensembles. - Create models for analyzing markets with MarketFlow. - Predict sporting events with SportFlow. - Develop trading systems and analyze portfolios using MarketFlow and Quantopian's pyfolio. Tech stack: Python Why it matters: AlphaPy is a powerful tool for anyone looking to automate trading systems and sports betting models using Python. It provides a comprehensive set of features and libraries to streamline the process of machine learning and data analysis.

### Overview: Learn LLM internals step by step - from tokenization to attention to inference optimization. ### Key features: - Tokenization - Attention - Inference optimization ### Tech stack: Python and Jupyter Notebooks ### Why it matters: This series provides a comprehensive guide to understanding the inner workings of Large Language Models (LLMs), from tokenization to attention and inference optimization, making it essential for developers looking to enhance their knowledge in this field.

Overview: GANDissect is a Pytorch-based tool designed to visualize and understand the neurons of a Generative Adversarial Network (GAN). It provides dissection results as both static summaries and interactive visualizations. Key features: - Dissects a GAN model - Offers static summary and interactive visualization options - Supports dissecting multiple layers and units within a GAN Tech stack: Python Why it matters: GANDissect simplifies the analysis of complex GAN models, making it easier to understand the internal representations and units of a GAN, which can aid in diagnosing and improving GAN performance.

### Overview: AIdea Server is a fully open-source application server built with Go, integrating GPT and large language models like Stable Diffusion and AI models for chat and image generation. ### Key Features: - Supports GPT and domestic language models (Stable Diffusion, SDXL1.0, etc.). - Integrates Stable Diffusion for image generation. - Compatible with OpenAI API and other AI models. ### Tech Stack: Go ### Why It Matters: AIdea Server provides a versatile platform for AI chat and image generation, making it useful for developers and creatives alike. Repository: mylxsw/aidea-server GitHub Stars: 1765

Overview: FSL-Mate is a comprehensive collection of resources for few-shot learning (FSL), including a paper list and a PaddlePaddle-based Python library for FSL. Key features: - FewShotPapers: A paper list tracking research advances in FSL. - PaddleFSL: A Python library for FSL using PaddlePaddle. Tech stack: Python Why it matters: FSL-Mate simplifies the process of researching and implementing few-shot learning, making it easier for developers to explore and apply FSL techniques.

Overview: Vocal Remover is a deep-learning-based tool designed to extract the instrumental track from your songs, enhancing the audio quality significantly. Key features: - Extracts instrumental tracks from audio files - Supports both CPU and GPU execution - Provides advanced options for improved separation quality - Includes a dataset for training your own model Tech stack: Python Why it matters: This tool is essential for audio engineers and enthusiasts looking to enhance their music production by removing vocals from instrumental tracks, thereby improving the overall sound quality and making the music more versatile. Repository: tsurumeso/vocal-remover Primary language: Python GitHub stars: 1764

Overview: Loam_velodyne is a real-time method for state estimation and mapping using a 3D lidar, designed for applications in robotics and autonomous systems. Key features: - Real-time state estimation and mapping - Uses a 3D lidar for accurate data collection - Compatible with ROS Indigo and Velodyne VLP16 Tech stack: C++ Why it matters: Loam_velodyne provides a robust and efficient solution for real-time mapping and localization in environments with 3D lidar sensors, making it valuable for various robotics and autonomous systems applications.

### Overview: E2B Runtime is the open-source backend powering E2B Cloud, the AI agent cloud, providing hardware-isolated sandboxes, pause and resume functionality, and more. ### Key features: - **Hardware-isolated sandboxes**: One Firecracker microVM per sandbox, with isolation in cgroup and network namespace. - **Pause and resume**: Sandboxes resume from a snapshot, and pause diffs memory and disk against the template. - **Fork a running sandbox**: Forks checkpoints a live sandbox in place and starts new sandboxes from that snapshot. - **Templates built from your recipe**: Layered builds from Docker images and build steps, each layer hashed and cached. ### Tech stack: Go ### Why it matters: E2B Runtime provides a hardware-isolated environment for agent sessions, enabling fast and reliable sandbox management, and is the foundation for the E2B AI agent cloud.

### Overview: MAgent is a platform designed for many-agent reinforcement learning research, focusing on scaling up from hundreds to millions of agents. It aims to support advanced reinforcement learning research that goes beyond traditional single-agent or small-agent setups. ### Key features: - **Many-Agent Reinforcement Learning Platform**: Supports reinforcement learning research with a large number of agents. - **AAAIArticle**: Demonstrates a research paper on MAgent. - **Demo Videos**: Shows interesting demonstrations of MAgent's capabilities. - **Rule-Based and Deep Learning**: Supports both rule-based algorithms and deep learning frameworks. ### Tech stack: Python ### Why it matters: MAgent is essential for researchers and developers looking to advance the field of many-agent reinforcement learning, providing a robust platform for scaling up their research efforts. Repository: geek-ai/MAgent Primary language: Python GitHub stars: 1762

Overview: VoteNet is a deep Hough voting-based 3D object detection network designed for point cloud data, achieving state-of-the-art performance on ScanNet and SUN RGB-D datasets. Key features: - End-to-end 3D object detection network - Synergy of deep point set networks and Hough voting - State-of-the-art 3D detection on real 3D scans - Simple design, compact model size, high efficiency Tech stack: Python, PyTorch, TensorFlow, CUDA 10.0, cuDNN v7.4 Why it matters: VoteNet leverages geometric information to directly detect objects in point clouds, overcoming challenges in sparse data representation and improving detection accuracy without relying on color images.

Overview: TFB is a comprehensive benchmarking project for time series forecasting methods, aiming to evaluate and compare various models across different datasets and scenarios. Key features: - Multiple datasets including PEMS03, PEMS07, and 27 multivariate datasets. - Support for predicting only a subset of input variables. - Fair benchmarking with diverse evaluation metrics and criteria. - Open-source code and documentation in both English and Chinese. Tech stack: Shell, Python, PyTorch. Why it matters: TFB provides a valuable resource for researchers and practitioners to evaluate and compare time series forecasting methods, ensuring fair and comprehensive benchmarking across different datasets and scenarios.