Projects

Overview: Mine AI is an open-source toolkit designed to advance AI-enabled payment-security, fraud-detection, and AML/CFT compliance for U.S.-accessible digital-asset and blockchain-based payment systems. It includes research and reference-implementation infrastructure for modern payment-security stacks. Key features: - Graph-Temporal Fraud Detection Models - LLM-Powered Interpretable Compliance Reasoning Tech stack: Python Why it matters: Mine AI provides reproducible and inspectable methods for better understanding and improving AI-enabled payment-security systems, addressing class imbalance, graph-structured nature, and interpretability requirements.

### Overview: Lagom is a PyTorch infrastructure designed for rapid prototyping of reinforcement learning algorithms. It balances flexibility and usability, providing modular tools for efficient experimentation without sacrificing flexibility. ### Key features: - Modular design for easy experimentation - Base classes for multiprocessing (master-worker framework) - Supports hyperparameter search via grid or random search - Built on PyTorch, suitable for reinforcement learning algorithms ### Tech stack: Python, PyTorch ### Why it matters: - Balances flexibility and usability - Supports efficient experimentation - Built on PyTorch, suitable for reinforcement learning ### Repository: - Repository: zuoxingdong/lagom - GitHub stars: 377 - GitHub description: lagom: A PyTorch infrastructure for rapid prototyping of reinforcement learning algorithms. - Primary language: Jupyter Notebook ### Table of Contents: - Installation - Documentation - RL Baselines - How to use lagom - Examples - Test - What's new ### Installation: - Recommended: Miniconda environment - Dependencies: Install required packages via `pip install -r requirements.txt` - From source: Clone repository and install via `pip install -e .` ### Documentation: - Available online at [http://lagom.readthedocs.io](http://lagom.readthedocs.io) ### RL Baselines: - Implemented standard RL algorithms ### How to use lagom: - Define RL agent - Define environment - Define engine for training and evaluation - Define configurations for hyperparameter search

### Overview: WikiPron is a command-line tool and Python API designed to mine multilingual pronunciation data from Wiktionary. It provides a comprehensive database of pronunciation pairs for over 5 million words across 348 languages, making it an invaluable resource for language researchers and developers. ### Key features: - **Command-line Tool**: A simple command-line interface for quick data scraping. - **Python API**: A robust Python API for integrating WikiPron functionalities into larger projects. - **Data**: Over 5 million word pronunciation pairs from 348 languages, including broad and narrow transcriptions. - **Models**: Supports various models for pronunciation mining and analysis. - **Development**: Easy installation and usage, with detailed documentation and examples. ### Tech stack: Python ### Why it matters: WikiPron is crucial for researchers and developers looking to study and analyze multilingual pronunciation data. Its comprehensive database and flexible API make it an indispensable tool in the field of language technology and linguistics. Repository: CUNY-CL/wikipron Primary language: Python GitHub stars: 375 GitHub description: Massively multilingual pronunciation mining

Overview: SpeechGPT 2.0-preview is a real-time spoken dialogue system designed for direct speech-to-speech interaction, featuring advanced capabilities in style generalization, role-playing, and versatile voice performance. It is built on a large language model trained on millions of hours of speech data, aiming to provide a seamless and interactive experience. Key features: - Style Generalization: Precise control over multiple emotions, styles, and voice types. - Role-Playing: Simulating various roles and their corresponding tones and expressions. - Multi-Modal Performance: Supports poetry recitation, storytelling, and dialects. - Real-Time Interaction: Offers smooth and responsive interaction for users. Tech stack: Python Why it matters: SpeechGPT 2.0-preview is a significant advancement in the field of spoken dialogue systems, offering a highly realistic and versatile interaction experience. Its ability to handle various styles and tones, combined with its real-time interaction capabilities, makes it a valuable tool for developers and researchers in the field. Repository: OpenMOSS/SpeechGPT-2.0-preview GitHub stars: 375 GitHub description: A real-time spoken dialogue system for direct speech-to-speech interaction. It is built on a large language model trained on millions of hours of speech data, aiming to provide a seamless and interactive experience. Read this in [English](./README_EN.md).

### Overview Digital Life Project 2 (DLP3D) is an open-source real-time framework that enables Large Language Models (LLMs) to bring expressive 3D avatars to life. Users converse naturally by voice, while characters respond on demand with unified audio, whole-body animation, and physics simulation directly in the browser. Characters are fully customizable in both appearance (3D models) and personality (character prompts) and can be easily adapted to any LLM or text-to-speech (TTS) service. ### Key Features - **Voice Interaction**: Users converse naturally through voice. - **Real-Time 3D Avatars**: Characters respond to speech with unified audio, whole-body animation, and physics simulation. - **Customization**: Characters can be fully customized in appearance and personality. - **Integration**: Easily adaptable to any LLM or TTS service. ### Tech Stack - TypeScript ### Why It Matters DLP3D provides a powerful and flexible solution for integrating LLMs with real-time 3D avatars, making it accessible and easy to use for developers and creators alike. Its open-source nature allows for easy integration and customization, making it a valuable tool for anyone looking to bring AI to life in their projects. ### GitHub Stars 374 ### Primary Language TypeScript ### Repository [dlp3d-ai/dlp3d.ai](https://github.com/dlp3d-ai/dlp3d.ai)

Overview: multi_task_NLP is a utility toolkit designed to simplify the training and inference of a single model for multiple natural language processing (NLP) tasks. This toolkit supports various data formats and multiple transformer-based encoders, making it an efficient solution for developers looking to build conversational AI systems. Key features: - Support for multiple NLU tasks - Single model for multiple tasks - Easy data preparation using YAML task files - Built-in transformer-based encoders (BERT, Distil-BERT, ALBERT, RoBERTa, XLNET) Tech stack: Python Why it matters: multi_task_NLP streamlines the process of building conversational AI systems by enabling developers to train a single model for multiple tasks, reducing resource consumption and latency, and making the system easier to manage.

Overview: Word2word is an easy-to-use Python package that provides high-quality bilingual word translations for 3,564 language pairs across 62 unique languages. Key features: - Access to top-k word translations for any given word in multiple languages - Ability to build a custom bilingual lexicon from a custom parallel corpus - Efficient and scalable approach for constructing large bilingual dictionaries Tech stack: Python Why it matters: Word2word simplifies the process of obtaining accurate translations for various languages, making it valuable for researchers, translators, and language enthusiasts looking to quickly access translations without the need for specialized knowledge or extensive resources. Repository: kakaobrain/word2word GitHub stars: 375

Overview: This project is a framework designed to benchmark different vector search engines under the same hardware constraints, allowing developers to choose the best engine for their needs. Key features: - Benchmarking of various vector search engines - Compatibility with different engines and datasets - Configuration and parameter management for both engines and datasets - Results are stored in the `./results/` directory Tech stack: Python Why it matters: This framework provides a standardized way to benchmark vector search engines, ensuring that developers can easily compare and choose the best engine for their specific requirements.

Overview: The SpeechBrain project is a comprehensive toolkit for speech processing, built entirely on PyTorch. It includes features such as speech recognition, speaker recognition, speech enhancement, and multi-microphone speech processing. Key features: - Speech recognition (HMM/DNN and end-to-end) - Speaker recognition - Speech enhancement - Speech separation - Multi-microphone speech processing Tech stack: PyTorch Why it matters: SpeechBrain simplifies the development of speech processing systems, making it accessible to developers and researchers alike.

Overview: This project provides a minimal codebase for finetuning large multimodal models, supporting various models like LLaVA-1.5, Qwen-VL-Chat, and Qwen2.5-VL-3B-Instruct. It abstracts the finetuning process and leverages the Hugging Face official implementation for seamless integration with existing models. Key features: - Support for finetuning various LMMs including single image, multiple/interleaved image, video, and unified models. - Abstracted components for easy integration and minimal effort. - Flexible and lightweight structure for easy understanding and modification. Tech stack: Python. Why it matters: This project simplifies the finetuning process of large multimodal models, making it accessible and easy to use for developers.

Overview: Floe Agent is an open-source AI agent workspace designed for iOS and Android, providing on-device Linux, terminal, code, Git, and local web previews capabilities. It supports BYOK cloud AI and includes features like Notes, Office documents, creative tools, development, voice conversations, and skills. Key features: - Real Linux environment on the device - Code, packages, and services run on the device - Supported local models for on-device inference - BYOK cloud AI using your own API keys Tech stack: Swift Why it matters: Floe Agent offers a comprehensive workspace for iOS and Android, combining AI tasks, notes, and development tools with a real Linux environment, making it a valuable tool for developers and AI enthusiasts. Repository: JiangNanGenius/floe-agent GitHub stars: 355 GitHub description: Open-source AI agent workspace for iOS/iPadOS and Android. On-device Linux, terminal, code, Git and local web previews; BYOK cloud AI. iOS adds Notes, Canvas, CAD/3D, voice, Skill Hub and supported local models. Android alpha includes Office viewing and editing. iPad-first.

### Overview: DataMate is an enterprise-level data processing platform designed for model fine-tuning and RAG retrieval, offering a comprehensive suite of core modules and visual orchestration tools. ### Key features: - **Data Collection**: Enables seamless data ingestion. - **Data Management**: Facilitates efficient data organization and storage. - **Operator Marketplace**: Provides a rich ecosystem of pre-built operators. - **Data Cleaning**: Enhances data quality through rigorous processing. - **Data Synthesis**: Facilitates data integration and transformation. - **Data Annotation**: Supports data labeling for improved accuracy. - **Data Evaluation**: Provides metrics and insights for data quality. - **Knowledge Generation**: Generates valuable knowledge from processed data. ### Tech stack: TypeScript ### Why it matters: DataMate is a powerful tool for enterprise-level data processing, making it an essential resource for model fine-tuning and RAG retrieval. Its comprehensive features and robust ecosystem make it a valuable asset for developers and data scientists alike.

### Overview: FLASH (Flash Attention Unit) is a novel implementation of the Transformer architecture proposed in the paper "Transformer Quality in Linear Time" by Weizhe Hua, Zihang Dai, Hanxiao Liu, and Quoc V. Le. This implementation aims to improve the efficiency and performance of the Transformer model by replacing multi-headed attention with a gated attention unit that can reduce the complexity to linear time. ### Key Features: - **Gated Attention Unit (GAU):** A novel attention mechanism that replaces multi-headed attention with a gated unit, reducing the computational complexity to linear time. - **ReLU Squared Activation:** Replaces the softmax activation with a ReLU squared activation, inspired by the Primer paper and ReLA Transformer. - **Grouped Linear Attention:** Combines GAU with Katharopoulos linear attention, using sequence grouping to address issues with autoregressive linear attention. - **FLASH Transformer:** A combination of GAU and linear attention, named after the "Flash" paper, which significantly improves performance and stability. - **Rotary Embedding (RoPE):** Rotary embedding is applied to both GAU and linear attention, improving the efficiency and performance of the model. ### Tech Stack: The project is implemented using Python, utilizing the FlashPytorch library. ### Why It Matters: The FLASH model, as described in the paper, offers a significant improvement in terms of computational efficiency and performance. By reducing the complexity of the Transformer model to linear time and addressing the issues with autoregressive linear attention, FLASH provides a more efficient and stable solution for natural language

Overview: SynapsCAD is an AI-powered 3D CAD IDE that allows developers to edit code, visualize designs in 3D, and interact with natural language to reshape their creations. It integrates OpenSCAD, Bevy, and Rust, offering a seamless experience for both code and visual editing. Key features: - Real-time code editing with OpenSCAD - Interactive 3D visualization - AI-assisted design suggestions - Support for multiple file formats (STL, OBJ, 3MF) - Compilation and preview capabilities Tech stack: Rust, OpenSCAD, Bevy, Hyperreal Why it matters: SynapsCAD simplifies the process of designing and visualizing complex 3D models, making it easier for developers to create, modify, and share their projects. Its integration of natural language and AI features enhances the design process, improving efficiency and creativity.

Overview: Curie is the first AI-agent framework designed for automated and rigorous scientific experimentation, enabling researchers to move research at the speed of thought. Key features: - 🚀 Automated Experimentation: From hypothesis formulation to result analysis and reflection. - 📊 Rigor Enhancement: Built-in verification modules enforce methodical procedures, agent reliability, and reproducibility. - 🔬 Broad Applicability: Supports ML Engineering, system analysis, and scientific discovery. - 💻 Use Starter Code: Supports working on arbitrary user's starter code. - 📂 Bring Your Own Dataset: Supports working on arbitrary user's datasets. - 🧾 Automatic, Insightful Reporting: Provides a sample report for a histopathologic cancer detection experiment. Tech stack: Python Why it matters: Curie empowers scientists to conduct rigorous and automated scientific experiments, ensuring precision, reliability, and reproducibility in their research. Repository: Just-Curieous/Curie Primary language: Python GitHub stars: 373 GitHub description: ❓Curie: Automated and Rigorous Scientific Experimentation with AI Agents README excerpt: # Curie: A Research Experimentation Agent <!-- # Curie: Automate Rigorous Scientific Experimentation --> [](https://arxiv.org/abs/2502.16069) [![License](https://img.shields.io/badge

### Overview: Codai is an open-source coding agent designed to enhance your terminal experience. It allows you to read, search, and edit code files directly within your project, with full control over your project's models and policies. ### Key Features: - **Plan, then Act**: Start in read-only mode, switch to Agent mode for implementation. - **Project-Wide Edits**: Edit real files with project guidance, ensuring context is preserved. - **On-Demand Context**: Open sessions without full-project indexing, discover files as needed. - **Controlled Execution**: Review writes and commands under your approval policy. - **Focused Terminal UI**: Scroll transcript with input pinned, expand action steps, command output, and file diffs. - **Your Choice of Models**: Connect to hosted providers or local servers, including Ollama. - **Reusable Workflows**: Apply project guidance, rules, skills, commands, and memory documents. - **Long-Running Work**: Compact context and resume a previous session with `codai code --continue`. - **Extensible Tools**: Add MCP servers, .NET plugins, and lifecycle hooks under the same approval policy. ### Tech Stack: C# ### Why It Matters: Codai simplifies coding by allowing you to work directly within your project files, enhancing productivity and control over your development process. Its ability to seamlessly integrate with various models and providers makes it a versatile tool for developers and coders alike.

### Overview The ***aleatory*** Python library provides functionality for simulating and visualizing stochastic processes. It introduces objects representing various stochastic processes and methods to generate realizations/trajectories and create visualizations. ### Key Features - **Arithmetic Brownian Motion** - **Bessel Process** - **Brownian Bridge** - **Brownian Excursion** - **Brownian Meander** - **Constant Elasticity Variance (CEV) process** - **Cox–Ingersoll–Ross (CIR) process** - **Chan-Karolyi-Longstaff-Sanders (CKLS) process** - **Fractional Brownian Motion process** - **Galton-Watson process with Poisson branching** - **Gamma process** - **Gaussian Process with Constant Kernel** - **Gaussian Process with Linear Kernel** - **Gaussian Process with Matern Kernel** - **Gaussian Process with Periodic Kernel** ### Tech Stack Python ### Why It Matters The ***aleatory*** library is essential for researchers, developers, and students working with stochastic processes in one dimension. It simplifies the simulation and visualization of complex stochastic models, making it easier to analyze and understand their behavior.

Overview: GraphAgent is an automated agent pipeline designed to address both explicit graph dependencies and implicit graph-enhanced semantic inter-dependencies, providing effective solutions for predictive and generative tasks on diverse datasets. Key features: - Graph Generator Agent - Task Planning Agent - Task Execution Agent Tech stack: Jupyter Notebook Why it matters: GraphAgent offers a comprehensive solution for integrating language models with graph language models, enabling the discovery of intricate relational information and data semantic dependencies, making it highly effective across various settings. Repository: HKUDS/GraphAgent Primary language: Jupyter Notebook GitHub stars: 372 GitHub description: [EMNLP2025] "GraphAgent: Agentic Graph Language Assistant"

### Overview: The repository contains a collection of resources and tools focused on Portuguese language development and analysis. This includes datasets such as multi-modal Portuguese captions, medical texts, news articles, and more. The datasets are designed to aid in various NLP tasks, including sentiment analysis, automatic translation, and aspect-based annotation. ### Key features: - Multi-modal Portuguese captions from Instagram posts. - 18th-century Portuguese medical texts. - Automatic translation of AG's corpus of news articles. - Stanford Alpaca dataset translated into Brazilian Portuguese. - Aspect-based annotated dataset of web consumer reviews. - Sequence of ASSIN for semantic similarity and entailment annotations. - Automated essay scoring dataset for Portuguese. - Aya Dataset PT. - Brazilian Court Decisions dataset. - Brazilian E-Commerce dataset. - Brazilian Headlines Sentiments dataset. - Brazilian Portuguese Literature Corpus. - Brazilian Portuguese Narrative Essays Dataset. - Brazilian Portuguese Sentiment Analysis Datasets. - Judgments of Federal Court of Accounts - Brazil (TCU). - A Large-Scale Dataset of Web Pages in the Portuguese Language. - Product reviews dataset. - A Corpus Geral do Português Brasileiro Contemporâneo. ### Tech stack: Python, Hugging Face, and various NLP libraries. ### Why it matters: These datasets and resources are crucial for researchers and developers working on Portuguese language NLP tasks. They provide a rich set of annotated data for training and evaluating models, ensuring that researchers can develop and test their NLP algorithms effectively.

Overview: This project provides a Python interface to the CMU Sphinxbase and Pocketsphinx libraries, enabling developers to integrate speech recognition capabilities into their projects. Key features: - LiveSpeech: An iterator class for continuous recognition or keyword search from a microphone. - AudioFile: An iterator class for continuous recognition or keyword search from a file. Tech stack: Python, SWIG, Setuptools, Sphinxbase, Pocketsphinx, CMU Sphinx, SWIG, Setuptools, Sphinxbase, Pocketsphinx, CMU Sphinx, SWIG, Setuptools, Sphinxbase, Pocketsphinx, CMU Sphinx, SWIG, Setuptools, Sphinxbase, Pocketsphinx, CMU Sphinx, SWIG, Setuptools, Sphinxbase, Pocketsphinx, CMU Sphinx, SWIG, Setuptools, Sphinxbase, Pocketsphinx, CMU Sphinx, SWIG, Setuptools, Sphinxbase, Pocketsphinx, CMU Sphinx, SWIG, Setuptools, Sphinxbase, Pocketsphinx, CMU Sphinx, SWIG, Setuptools, Sphinxbase, Pocketsphinx, CMU Sphinx, SWIG, Setuptools, Sphinxbase, Pocketsphinx, CMU Sphinx, SWIG, Setuptools, Sphinxbase, Pocketsphinx, CMU Sphinx, SWIG, Setuptools, Sphinxbase, Pocketsphinx, CMU Sphinx, SWIG, Setuptools, Sphinxbase, Pocketsphinx, CMU Sphinx, SWIG, Setuptools, Sphinxbase, Pocketsphinx, CMU Sphinx, SWIG, Setuptools, Sphinxbase,