Projects

### Overview: Exam Cram Coach is an AI exam-prep tutor designed for Claude Code, Codex, Cursor, Windsurf, Antigravity, and Gemini CLI. It integrates lecture slides, notes, homework, and past papers to provide a comprehensive and accurate review. The tool ensures that every concept is cited from specific files, prevents fabrication, and tracks mistakes for spaced re-education. ### Key Features: - **Cites Exact References**: Every concept is backed by specific file citations, preventing fabrication. - **Persistent Memory**: Maintains state locally, ensuring state is preserved across sessions. - **Figure Cropping**: Automatically extracts figures from lecture notes and homework, ensuring accurate presentations. - **Quiz Generation**: Quizzes are generated from real homework and exam questions, providing a realistic practice environment. - **Mistake Tracking**: Tracks and prioritizes mistakes for spaced re-education. - **Comprehensive Review**: Teaches chapters in plain language, quizzes with real questions, and compiles a concise cheat sheet. ### Tech Stack: - Python ### Why It Matters: Exam Cram Coach is essential for students preparing for exams, offering a reliable and accurate review system that ensures the material is cited correctly and provides a realistic practice environment. Its ability to track and correct mistakes makes it an invaluable tool for improving understanding and performance. ### Repository: Repository: ZeKaiNie/universal-examprep-skill Stars: 355 Description: Exam Cram Coach · 跨会话记忆与讲义溯源防幻觉的极速备考教练 | AI exam-prep tutor

Overview: This project is a PyTorch implementation of OpenAI GPT-2, providing model training, sentence generation, and metrics visualization. It is designed to be understandable and optimized. Key features: - Model training - Sentence generation - Metrics visualization Tech stack: Python Why it matters: This implementation is both understandable and optimized, making it a valuable resource for developers looking to work with language models. Repository: affjljoo3581/GPT2 Primary language: Python GitHub stars: 364 GitHub description: PyTorch Implementation of OpenAI GPT-2

Overview: This project is an efficient, scalable, and user-friendly MLX-based vision and language model API server that provides an OpenAI-compatible interface. It is built using Python and the FastAPI framework, making it ideal for local deployment on Apple Silicon with macOS. Key features: - Supports text, multimodal, image, embedding, and Whisper model endpoints. - Provides an OpenAI-compatible interface for MLX models. - Developed using Python and FastAPI. - Efficient, scalable, and user-friendly solution. Tech stack: Python and FastAPI. Why it matters: This project offers a practical solution for developers looking to run MLX-based models locally with an OpenAI-compatible interface, making it easier to integrate and deploy models without the need for expensive cloud services.

### Overview: Spanish word embeddings are available for download, computed using various methods and corpora. The embeddings are provided in vector and binary formats, along with statistics and descriptions of the used parameters and corpora. ### Key features: - **FastText embeddings from Spanish Unannotated Corpora**: Computed using FastText with Skipgram, this corpus contains 1,313,423 vectors with 300 dimensions. The embeddings were created using parameters such as min subword-ngram, max subword-ngram, and epochs. - **FastText embeddings from Spanish Billion Word Corpus**: This corpus contains 855,380 vectors with 300 dimensions. The embeddings were computed using FastText with Skipgram, with parameters including min subword-ngram, max subword-ngram, and epochs. - **Glove embeddings from Spanish Billion Word Corpus**: This corpus contains 855,380 vectors with 300 dimensions. The embeddings were computed using GloVe, with parameters including min subword-ngram, max subword-ngram, and epochs. - **Word2Vec embeddings from Spanish Billion Word Corpus**: This corpus contains 1,000,653 vectors with 300 dimensions. The embeddings were computed using Word2Vec, with parameters including min subword-ngram, max subword-ngram, and epochs. ### Tech stack: - **Embedding Formats**: Vector (.vec) and Binary (.bin) formats. ### Why it matters: These embeddings can be used for

### Overview: Expert Advisor Studio is a powerful tool for creating expert advisors and automating trading on web browsers. It includes features such as a FIX API gateway, a CRM, and a market maker bot, making it a versatile solution for financial trading. ### Key Features: - **Expert Advisor Generator**: Quickly create and test trading strategies. - **FIX API Gateway**: Connect with brokers and liquidity providers via the FIX protocol. - **CRM Integration**: Seamless integration with trading platforms for better management. - **Market Maker Bot**: Automated market making capabilities for better liquidity. ### Tech Stack: JavaScript ### Why It Matters: Expert Advisor Studio simplifies the process of creating and managing trading strategies, making it accessible to both beginners and experienced traders. Its robust features and integration capabilities make it a valuable tool for financial trading and automation.

### Overview: This repository is a curated list of AI art courses, tools, libraries, people, and places. It serves as a comprehensive resource for developers interested in learning about AI and its applications in the arts. ### Key Features: - Courses - Tools - People - Places ### Tech Stack: Python, JavaScript, Markdown, and GitHub Pages. ### Why It Matters: This list is essential for developers looking to integrate AI into their creative projects, providing a solid foundation in both AI and art. ### Repository: Repository: jonathandinu/ai4artists GitHub stars: 364 ### Primary Language: Unknown ### GitHub Description: A list of AI Art courses, tools, libraries, people, and places. Mainly tools and tutorials but also with some inspiring people and places thrown in too!

Overview: This package is designed for Chinese text segmentation, keyword extraction, and speech tagging. Key features: - Chinese text segmentation - Keyword extraction - Speech tagging Tech stack: R Why it matters: This package offers powerful tools for text processing, making it valuable for applications requiring natural language processing tasks in Chinese.

Overview: diff-sampler is an open-source toolbox designed for fast sampling of diffusion-based generative models. It enables fair comparisons among existing approaches and helps researchers develop better algorithms. Key features: - Supports a wide range of model implementations, numerical solvers, time schedules, and advanced features. - Official implementations of our works published in ICML'24, NeurIPS'24, CVPR'24, and J. Stat. Mech'25. Tech stack: Jupyter Notebook Why it matters: diff-sampler provides a platform for researchers to evaluate and compare diffusion-based generative models, facilitating the development of more efficient and accurate algorithms. Repository: zju-pi/diff-sampler GitHub stars: 364

Overview: Melusine is a comprehensive email processing library designed to optimize your email workflow. It offers advanced features such as effortless email routing, smart prioritization, snippet summaries, and precision filtering. Key features: - Seamless email routing - Efficient email prioritization - Precise email snippet extraction - Robust email filtering Tech stack: Python Why it matters: Melusine simplifies the process of email processing, making it easier to manage and analyze emails, while also ensuring that your workflow is optimized and efficient.

Overview: `cargo-spellcheck` is a Rust crate designed to check your documentation for spelling and grammar mistakes using `hunspell` and `nlprule` for grammar. It helps streamline the review process and improve CI checks. Key features: - Checks spelling and grammar in your documentation - Fixes typos automatically - Integrates with CI/CD systems - Supports both `hunspell` and `nlprule` for different needs Tech stack: Rust Why it matters: It simplifies the review process and improves the quality of your documentation, making it easier to maintain and improve over time. Repository: drahnr/cargo-spellcheck GitHub stars: 363 GitHub description: Checks all your documentation for spelling and grammar mistakes with hunspell and a nlprule based checker for grammar

### Overview: The **LangChain Documentation Helper** is an intelligent AI-powered web application designed to assist developers in understanding and utilizing LangChain documentation. This tool leverages advanced Retrieval-Augmented Generation (RAG) techniques, enhanced with web crawling capabilities and conversational memory, to provide accurate answers to questions about LangChain documentation. ### Key Features: - 🌐 **Web Crawling**: Real-time web scraping and content extraction using Tavily's advanced crawling capabilities. - 📚 **Document Processing**: Intelligent chunking and preprocessing of LangChain documentation. - 🔍 **Vector Storage**: Advanced embedding and indexing using Pinecone for fast similarity search. - 🎯 **Intelligent Retrieval**: Context-aware document retrieval based on user queries. - 🧩 **Memory System**: Conversational memory for coreference resolution and context continuity. - 🧠 **Context-Aware Generation**: Provides accurate, contextual answers with source citations. - 💬 **Interactive Interface**: User-friendly chat interface powered by Streamlit. - 🚀 **Real-time Processing**: Fast end-to-end pipeline from query to response. ### Tech Stack: - 🖥️ **Frontend**: Streamlit - 🧠 **AI Framework**: LangChain 🦜🔗 - 🔍 **Vector Database**: Pinecone 🌲 - 🌐 **Web Crawling**: Tavily 🌍 - 🧩 **Memory**: Conversational Memory - 🐍 **Backend**: Python - 🐍 **Backend**: Python ### Why It Matters: This application

### Overview: This project provides a customizable pipeline for extracting and preprocessing multimodal data from the MIMIC-IV dataset, enabling researchers to prepare and analyze patient data for various downstream tasks. ### Key features: - **Data Cleaning**: Removes outliers and imputes missing entries. - **Clinical Grouping**: Allows users to group medical features using standard coding systems for dimensionality reduction. - **Customization**: Users can generate a personalized patient cohort by recording customization steps. - **Time-Series Binning**: Produces a smooth time-series dataset by binning sequential data into equal-length time intervals. - **Modular Models**: Includes several commonly used sequential models for prediction tasks. - **Model Evaluation**: Provides standard methods for assessing model performance and options for reporting individual and group fairness measures. ### Tech stack: Jupyter Notebook ### Why it matters: This pipeline simplifies the process of preparing and analyzing patient data from the MIMIC-IV dataset, making it easier for researchers to identify patterns and trends in patient health.

Overview: SwiftLLM is a tiny yet powerful LLM inference system designed specifically for researching purposes. It achieves equivalent performance to vLLM with only 2k lines of code, making it easy to understand, modify, and integrate with your research ideas. Key features: - Iterational Scheduling and Selective Batching - PagedAttention - LLaMA models and variants - Piggybacking Prefill and Decoding - Flash Attention - Paged Attention v2 Tech stack: Python and OpenAI Triton Why it matters: SwiftLLM is tailored for researching purposes, providing a compact and easy-to-modify solution without compromising performance, making it ideal for researchers looking to integrate their innovative ideas into their projects. Repository: interestingLSY/swiftLLM GitHub stars: 361

Overview: RepoGPT is an AI-powered GitHub assistant that simplifies codebase exploration and management, providing intelligent insights to enhance development efficiency. Key features: - AI-driven repo interaction for chat and codebase management - Streamlined repository management with import, list, and delete capabilities - PostgreSQL + pgvector integration for efficient AI data processing and storage - Simple setup for easy deployment Tech stack: TypeScript Why it matters: RepoGPT revolutionizes how developers interact with their GitHub repositories, making codebase exploration and management more intuitive and efficient.

### Overview: DynaSaur is a dynamic Large Language Model (LLM)-based agent framework introduced at COLM 2025. It leverages a programming language as a universal representation for actions, enabling the creation of new actions from existing ones or from scratch. DynaSaur demonstrates remarkable versatility, recovering automatically when no relevant actions are available or when existing actions fail due to unforeseen edge cases. It holds the top position on the GAIA benchmark and remains the leading non-ensemble method to date. ### Key Features: - **Dynamic Action Generation**: DynaSaur generates Python snippets that either call on existing actions or create new ones based on the current action set. - **Reusable Library**: It gradually expands a reusable library of actions, making it easier to develop and reuse actions. - **Versatility**: The framework automatically recovers when no relevant actions are available or when existing actions fail due to unforeseen edge cases. ### Tech Stack: - **Python**: The primary programming language used to implement the framework. ### Why It Matters: DynaSaur is a significant advancement in the field of LLM-based agents, offering a dynamic and versatile approach to action generation that significantly outperforms traditional methods on benchmark tests.

### Overview: This repository is a curated list of awesome works related to high-dimensional structure/vector search and databases. It includes various tools and services for vector database management, vector search engines, and more. ### Key Features: - Google Vector Search (Vertex AI) - Pinecone - Weaviate - Vespa - txtai - marqo - vectara - Epsilla - algolia - Meilisearch - nucliadb - OpenSearch - MyScale - QdrantCloud - zilliz - OpenSearch's AlibabaCloud - Typesense's Cloud - MongoDB Atlas Vector Search - SuperDuperDB - KBD.AI - Denser Retriever - Rivestack - Managed PostgreSQL with pgvector for AI workloads - HNSW indexing - pgvector - Chroma - LlamaIndex - Epsilla - jvector - RAFT - Vald - Voyager - tinyvector - USearch - vearch - infinity - havenask - chromem-go - OasysDB - Meilisearch - arroy - bleve - cuVS - vsag - sqlite-vec ### Tech Stack: The repository uses a variety of technologies including Python, JavaScript, Rust, Go, and others. ### Why it Matters: This list serves as a comprehensive resource for developers looking to explore and implement vector database solutions, providing a curated collection of

Overview: This project, "COVID-QA," is an open-source initiative aimed at improving existing question answering and search methods by collecting and evaluating a Question Answering dataset. It also provides a web application that matches questions about COVID-19 with trustworthy answers using Natural Language Processing (NLP). Key features: - A Question Answering dataset for improving QA/search methods. - Question matching capabilities to provide accurate answers to questions about COVID-19. - A web application for users to ask questions and receive answers. Tech stack: The project utilizes Jupyter Notebooks as the primary language and employs the Haystack library for NLP tasks such as detecting similar questions in FAQs and extracting answers from free texts. Why it matters: The project addresses the challenge of finding reliable answers to frequently asked questions about the COVID-19 pandemic, which can be difficult due to the rapid evolution of information and the potential for outdated answers. By providing a structured and searchable dataset and a user-friendly interface, it aims to streamline the process of obtaining accurate and up-to-date information about the virus.

### Overview: Zeroth is an open-source project for Korean speech recognition using the Kaldi toolkit. It aims to make speech recognition technology more accessible by providing an official Korean Kaldi recipe. ### Key features: - Official Korean Kaldi recipe - Provides a starting point for building new products and services using speech recognition - Offers a voice recording app for contributing to the open-source database of Korean training data - Uses end-to-end data-driven language models and phonetic dictionaries - Supports TDNN, TDNN + LSTM, and TDNN + OPGRU acoustic models - Data augmentation of reverberant speech ### Tech stack: Shell ### Why it matters: Zeroth aims to democratize speech recognition technology by providing an accessible and open-source Korean recipe, making it easier for enterprises to integrate speech recognition into their B2C communications. ### Repository: goodatlas/zeroth ### GitHub stars: 362 ### Special thanks: Developed in collaboration with Wonkyum Lee at Gridspace Inc.

Overview: MatrixHub is an open-source, self-hosted AI model registry designed for large-scale enterprise inference, replacing Hugging Face for accelerated vLLM and SGLang performance. Key features: - Self-hosted AI model registry - Compatible with Hugging Face - Accelerates vLLM and SGLang workloads - Live demo available at demo.matrixhub.ai Tech stack: Go Why it matters: MatrixHub provides a reliable, self-hosted alternative to Hugging Face, speeding up inference times for large language models and scripting languages. Repository: matrixhub-ai/matrixhub GitHub stars: 359 GitHub description: An open-source, self-hosted AI model hub for large-scale enterprise inference.

Overview: This project trains a computer program to identify a voice as male or female using acoustic properties of the voice and speech analysis. Key features: - Voice Gender Recognition - Pre-processed dataset for training - Accuracy of 100% on training set and 89% on test set - Updated model with frequency range 0hz-280hz achieving 100%/99% accuracy Tech stack: R Why it matters: This project demonstrates the effectiveness of using machine learning algorithms to accurately identify gender based on voice characteristics, which has applications in fields such as voice recognition and biometric authentication.