Projects

### Overview: The AI plugin provides a set of opt-in AI features for WordPress, enabling authors, editors, and admins to test and experiment with AI capabilities directly within the platform. This modular framework allows developers, agencies, and hosts to build or extend AI-powered workflows using building blocks from the WordPress AI team. ### Key Features: - Image Generation and Editing - Text Generation - Voice Recognition - Chatbots ### Tech Stack: PHP ### Why it Matters: The AI plugin serves as a reference implementation for developers, agencies, and hosts looking to integrate AI capabilities into their WordPress sites. It provides a unified experience for testing and experimenting with AI features, making it an essential tool for anyone interested in leveraging AI in WordPress.

### Overview: This repository contains an implementation of the Enhanced Sequential Inference Model (ESIM) for natural language inference using PyTorch. The model was originally presented in the paper "Enhanced LSTM for Natural Language Inference" by Chen et al. in 2016. ### Key features: - ESIM model for natural language inference - PyTorch implementation - High-level architecture illustrated in the provided figure - Supports SNLI, GloVe embeddings, and MultiNLI datasets - Preprocessing scripts for SNLI, MultiNLI, and BNLI datasets - Training and testing scripts for the ESIM model ### Tech stack: PyTorch ### Why it matters: This implementation provides a practical tool for researchers and developers interested in natural language inference tasks, allowing them to quickly test and validate their models on various datasets and configurations.

Overview: This repository contains tutorials for the Haystack library, a powerful tool for natural language processing and information retrieval. The tutorials cover various aspects of using Haystack, including building QA pipelines, generating structured outputs, serializing pipelines, preprocessing different file types, filtering metadata, classifying documents and queries by language, and more. Key features: - Building QA pipelines with retrieval augmentation - Generating structured outputs using loop-based auto-correction - Serializing pipelines for easier maintenance and reusability - Preprocessing different file types for indexing - Metadata filtering for better query relevance - Classifying documents and queries by language for improved accuracy Tech stack: Jupyter Notebook Why it matters: These tutorials provide a comprehensive guide to using Haystack, making it easier for developers to build, maintain, and improve their information retrieval systems.

### Overview: MultiMed is a comprehensive series of multilingual multitask medical speech processing projects, including datasets, benchmarks, and pre-trained models for automatic speech recognition and spoken named entity recognition. The series is designed to advance research in healthcare applications through large-scale datasets and fine-tuned models. ### Key Features: - **VietMed Dataset:** A Vietnamese speech recognition dataset in the medical domain with 16 hours of labeled medical speech, 1000 hours of unlabeled medical speech, and 1200 hours of unlabeled general-domain speech. - **VietMed-NER Dataset:** A spoken named entity recognition dataset in the medical domain with 18 distinct types of entities. - **Pre-trained Models:** Release of pre-trained models for Vietnamese ASR and spoken NER, including w2v2-Viet, XLSR-53-Viet, and fine-tuned models for medical ASR. - **Publicly Available:** All code, data, and models are made publicly available on GitHub. ### Tech Stack: - **Python:** The primary programming language used for implementing and managing the projects. ### Why It Matters: MultiMed provides a valuable resource for researchers and practitioners in the field of healthcare technology, offering large-scale datasets and pre-trained models that can significantly advance the development of medical speech processing technologies.

Overview: This project aims to create a comprehensive voice toolkit for training, testing, and deploying speaker verification systems. Key features: - Speaker Verification System - Real-time Speaker Verification - Multi-Modal Speaker Verification Tech stack: Python, PyTorch, Long-Yuan Why it matters: This toolkit provides a robust solution for speaker verification, enabling developers to quickly integrate advanced speaker verification capabilities into their projects.

### Overview: Proof of Thought is an LLM-based reasoning tool that leverages Z3 theorem proving for advanced logical reasoning tasks. It supports multiple backend options, including SMT2 and JSON execution, and integrates seamlessly with Azure OpenAI for enhanced performance. ### Key Features: - **Dual Backend Support**: Choose between SMT2 (default) or JSON execution backends. - **Azure Integration**: Native support for Azure GPT-4o and GPT-5 models. - **Comprehensive Benchmarks**: Evaluated on 5 reasoning datasets (ProntoQA, FOLIO, ProofWriter, ConditionalQA, StrategyQA). - **High-level API**: Simple Python interface for reasoning tasks. - **Batch Evaluation Pipeline**: Built-in tools for dataset evaluation and metrics. - **Postprocessing Techniques**: Self-Refine, Self-Consistency, Decomposed Prompting, and Least-to-Most Prompting for enhanced reasoning quality. ### Tech Stack: - Python 3.12 - Z3 solver (automatically installed via `z3-solver` package) - OpenAI API key or Azure OpenAI endpoint - Z3 theorem proving ### Why it Matters: Proof of Thought provides a powerful and flexible way to perform complex reasoning tasks using LLMs and Z3 theorem proving, making it suitable for a wide range of applications in AI and automated reasoning.

Overview: This repository provides the source code and data for Tabular Transformers for Modeling Multivariate Time Series, a paper presented at ICASSP 2021. It includes modules for hierarchical transformers and a synthetic credit card transaction dataset. Key features: - Modules for hierarchical transformers for tabular data - Synthetic credit card transaction dataset - Modified Adaptive Softmax for handling masking - Modified DataCollatorForLanguageModeling for tabular data - Built within transformers from HuggingFace Tech stack: Python, Pytorch, HuggingFace / Transformer, scikit-learn, Pandas Why it matters: This repository offers practical tools and data for training tabular transformers, making it valuable for researchers and practitioners interested in time series modeling.

### Overview: ReinFlow is a project that fine-tunes flow matching policies using online reinforcement learning. The paper "ReinFlow: Fine-tuning Flow Matching Policy with Online Reinforcement Learning" was accepted at NeurIPS 2025. This project supports fine-tuning of VLA models like $\pi_0$ and $\pi_{0.5}$, and is fully open-sourced. ### Key Features: - **Fine-tuning Flow Matching Policies:** ReinFlow fine-tunes flow matching policies using online reinforcement learning. - **Support for VLA Models:** The project supports fine-tuning of VLA models such as $\pi_0$ and $\pi_{0.5}$. - **Fully Open-Sourced:** The code and checkpoints are available for download and are fully open-sourced. ### Tech Stack: - Python ### Why It Matters: ReinFlow is a significant contribution to the field of reinforcement learning and vision-language-action models, offering a practical solution for fine-tuning flow matching policies. The project's open-source nature ensures that researchers and practitioners can benefit from the code and results, making it a valuable resource for the community.

Overview: SlideDeck AI is an AI-powered tool designed to assist developers in creating professional PowerPoint slide decks for presentations. It leverages cutting-edge AI and large language models to streamline the process, making it easy to co-create stunning slide decks on any topic. Key features: - AI Content Generation: Generates initial slide content using a pre-defined schema. - Visual Enhancement: Uses keywords from the JSON output to download relevant images. - PPTX Assembly: Uses the `python-pptx` library to create slides based on structured JSON data. - Refinement & Iteration: Allows users to refine content and provide additional instructions. - Instant Download: Provides a download button for easy saving of the generated PowerPoint presentation. Tech stack: Python Why it matters: SlideDeck AI streamlines the creation process, making it easier and more efficient for developers to create professional slide decks, ensuring presentations are both visually appealing and well-organized.

Overview: XLingEval is a framework designed to evaluate Large Language Models (LLMs) for their effectiveness in healthcare queries across multiple languages. It focuses on three key criteria: correctness, consistency, and verifiability, using extensive experiments on four major global languages: English, Spanish, Chinese, and Hindi. Key features: - **XLingHealth**: A cross-lingual benchmark for examining LLMs' multilingual capabilities in healthcare. - **XLingEval**: An empirically-derived framework that evaluates LLM responses in terms of correctness, consistency, and verifiability. - **Algorithmic and Human Evaluation**: Combines algorithmic metrics with expert human evaluation strategies. Tech stack: Python Why it matters: XLingEval addresses the need for enhanced cross-lingual capabilities in LLMs, ensuring that these systems perform well in non-English languages, thereby promoting equity in their real-world use.

Overview: This project implements a Sudoku puzzle solver using the Tree-of-Thought (ToT) framework, a novel approach inspired by human problem-solving methods. The framework allows LLMs to explore solution paths through a multi-round conversation, enabling backtracking for improved accuracy. Key features: - Sudoku puzzle solver - Tree-of-Thought (ToT) framework for problem-solving - Augmented LLM with additional modules (prompter, checker, memory, ToT controller) Tech stack: Python Why it matters: This project showcases an innovative method to enhance the problem-solving capabilities of large language models, potentially improving their performance in complex reasoning tasks.

Overview: UTMOSv2 is an official implementation of the T05 system for the VoiceMOS Challenge 2024, showcasing state-of-the-art performance in naturalness MOS prediction of high-quality synthetic speech. Key features: - Accurate MOS prediction - Transfer learning from a deep image classifier - Pretrained model for quick predictions Tech stack: Python Why it matters: UTMOSv2 achieved 1st place in 7 out of 16 metrics and 2nd place in the remaining 9 metrics in the VoiceMOS Challenge 2024 Track1, demonstrating its effectiveness and reliability in naturalness MOS prediction. Repository: sarulab-speech/UTMOSv2 Primary language: Python GitHub stars: 376

### Overview: The **Awesome Jev Live** repository is a comprehensive index of evidence-graded resources for the TypeSafe System One ecosystem, including SDKs, MCP tools, agents, apps, and open models. It is rebuilt every two hours, featuring 801 entries in 20 languages, with 26 implementation languages. ### Key Features: - **Evidence-Graded Index**: The repository provides an evidence-graded index of resources for the TypeSafe System One ecosystem. - **Rebuild Frequency**: Entries are rebuilt every two hours, ensuring up-to-date information. - **Language Support**: The repository covers 20 languages and 26 implementation languages. - **Swift SDK**: The repository includes a Swift SDK for running chat, vision, and speech models on iPhone and Mac with Apple's Core AI. - **JavaScript SDK**: The repository features a JavaScript SDK for route optimization using the jev-router. - **Non-Autoregressive Decision Engine**: The repository includes a non-autoregressive System 1 decision engine for typed choice, score, and yes/no decisions. ### Tech Stack: The repository uses Python as the primary language. ### Why It Matters: The repository is essential for developers looking to integrate evidence-graded resources into their projects, ensuring that they have access to the latest and most relevant tools and models. It serves as a valuable resource for anyone working with the TypeSafe System One ecosystem, providing a reliable and up-to-date index of resources. ### GitHub Stars: The repository has received 376 GitHub stars, indicating a high level of interest

### Overview: MoE-Infinity is a PyTorch library designed for efficient, fast, and easy serving of Mixture-of-Experts (MoE) models on memory-constrained GPUs. It offloads expert weights to host memory and SSD, utilizing an activation-aware cache and prefetching to minimize transfer costs. ### Key Features: - **Cost-effective.** Expert offloading to host memory and SSD allows serving large MoE models that would otherwise not fit on memory-constrained GPUs. - **Fast.** Activation-aware expert caching, prefetching, tracing, fused CUDA kernels, Marlin INT4 GEMM, and FP4/MXFP4 expert paths keep the hot path lean. - **HuggingFace-native.** The `MoE` class remains the current in-process synchronous API, but `MoE.generate()` emits `DeprecationWarning` and is scheduled for removal. Use `MoE.serve()` for continuous batching; it starts an async HTTP service and is not a drop-in in-process return API. - **Production serving.** OpenAI-compatible HTTP server with continuous batching, paged KV cache, request scheduling with preemption, streaming (SSE), runtime hot reload, watchdog/health monitoring, and crash-recovery logging. Opt-in correctness-preserving prefix KV reuse (`--enable-prefix-caching`, default off) reuses exact shared prefixes on the supported Qwen3 + FlashInfer path with cold/warm equivalence; unsupported runtimes fall back to the cold path unchanged. - **Acceleration-aware.** Automatically integrates with [FlashAttention](https

Overview: This project details the process of building a large language model from scratch using PyTorch, inspired by the methods used in creating large-scale language models like those behind ChatGPT. Key features: - Step-by-step guide for creating a language model from scratch - Explanation of LLMs and their training methods - Use of Python and PyTorch for implementation Tech stack: Python, PyTorch Why it matters: This book provides a practical, step-by-step approach to understanding and building language models from the ground up, mirroring the methods used in large-scale models like ChatGPT. Repository: JohnMachado11/Build-a-Large-Language-Model-from-Scratch GitHub stars: 376 GitHub description: Building a GPT-like LLM from scratch with PyTorch.

### Overview: 将古诗词转化为竖屏中国画动态视频，保留原诗意境与环境声，通过 Docker 调用 Gemini I2V 实现逐句或两句一景的动态效果。 ### Key features: - 逐句或两句一景的竖屏中国画动态视频 - 保留原诗意境与环境声 - 保留原诗意境与环境声 - 保留原诗意境与环境声 - 保留原诗意境与环境声 - 保留原诗意境与环境声 ### Tech stack: Shell ### Why it matters: Skill 通过识别古诗词的意境和环境声，将静态的古画转化为动态的中国风视频，保留了原诗的意境和环境声，使得视频更具观赏性和艺术性。

Overview: This repository contains my work for the Coursera Natural Language Processing Specialization offered by deeplearning.ai. The code base, quiz questions, and diagrams are based on the specialization's course content. Key features: - Sentiment Analysis with Logistic Regression - Natural Language Preprocessing - Advanced Sentiment Analysis with Dense and Recurrent Neural Networks - Text Generation and Named Entity Recognition - Machine Translation and Question-Answering Tech stack: Jupyter Notebook Why it matters: This specialization equips learners with state-of-the-art deep learning techniques for building cutting-edge NLP systems, making it essential for professionals interested in AI-powered language analysis and AI development.

### Overview: EVE Series: Encoder-Free Vision-Language Models from BAAI ### Key features: - **Encoder-Free Vision-Language Models (EVE)**: Developed by BAAI, these models are designed to remove the vision encoder from Vision-Language Models (VLMs), achieving superior capability with arbitrary image aspect ratios. - **Data Efficiency**: Leverages only 100M publicly available data from OpenImages, SAM, LAION, and Datacomp for pre-training, making it highly efficient. - **Pioneering Route**: Provides an efficient, transparent, and practical training strategy for developing pure decoder-only architectures across modalities. ### Tech stack: Python ### Why it matters: EVE Series represents a significant advancement in the field of vision-language models, offering a practical and efficient solution for developing pure decoder-only architectures, thereby bridging the performance gap between encoder-free and encoder-based VLMs.

### Overview: Sentiment analysis is a crucial task in natural language processing, where the goal is to determine the emotional tone of a piece of text. This project introduces a neural network trained on the Stanford Sentiment Treebank, fine-tuned on BERT, ALBERT, or DistilBERT, to perform sentiment analysis efficiently. ### Key features: - Fine-tuning BERT, ALBERT, or DistilBERT on the Stanford Sentiment Treebank - Easy setup using Python 3.9 - Virtual environment for Python 3.9 - Installation of necessary packages - Testing with `pytest` - Running the server for sentiment analysis - Training your own model and using Sentiment Analysis with it ### Tech stack: Python ### Why it matters: This project provides a robust and efficient sentiment analysis tool, ideal for developers looking to automate sentiment analysis in their projects.

Overview: OasysDB is an in-memory vector store designed for efficient semantic caching and retrieval, utilizing Redis for semantic caching. Key features: - In-memory vector store - High-performance read and write performance - Supports semantic caching and retrieval Tech stack: Rust Why it matters: OasysDB offers a fast and efficient solution for semantic caching and retrieval, making it suitable for applications requiring quick access to large amounts of data. Repository: edwinkys/oasysdb GitHub stars: 376