Projects

Overview: Vicinity is a lightweight, low-dependency vector store designed for efficient nearest neighbor search. It provides a unified interface for various backends, making it easy to experiment with different indexing methods and distance metrics without needing to learn new interfaces. Key features: - Supports multiple backends: Basic, Memory, KNN, and more. - Flexible evaluation metrics: Cosine, Euclidean, and others. - Simple and intuitive interface for easy experimentation. - Easy installation and usage with `pip install vicinity`. Tech stack: Python Why it matters: Vicinity simplifies the process of comparing and evaluating different vector store backends, making it easier to find the best fit for any use case. It provides a straightforward way to measure performance, allowing developers to quickly identify the most effective method for their needs.

Overview: stt-ru is a comprehensive speech-to-text library that includes Python, CLI, web interface with REST API, and Docker container based on the GigaAM-v3 model. It provides accurate transcription of Russian speech, including punctuations and numbers, and supports stereo channels. Key features: - Python library for transcription - CLI tool for real-time transcription - Web interface with REST API for easy integration - Docker container for GPU and CPU compatibility - Model available on Hugging Face for downloading Tech stack: Python, CLI, web interface, Docker, Hugging Face model Why it matters: stt-ru simplifies the process of converting audio recordings into text, making it easier for developers to integrate speech recognition into their projects without needing to manage complex models or infrastructure.

### Overview: RefAlign is an innovative training-time alignment framework designed specifically for improving reference fidelity, identity consistency, and text controllability in reference-to-video generation. This project was presented at ECCV 2026 and has been widely recognized for its effectiveness in enhancing the quality and coherence of generated videos. ### Key Features: - **Representation Alignment**: RefAlign explicitly aligns DiT reference-branch features to the feature space of a frozen visual foundation model (VFM) during training. - **Improved Reference Fidelity**: By addressing copy-paste artifacts and multi-subject confusion, RefAlign enhances the fidelity of references in generated videos. - **Identity Consistency**: RefAlign ensures consistent identity across different subjects in the generated video, improving the overall quality. - **Stronger Semantic Discrimination**: The framework effectively discriminates between different semantic elements in the reference and video, leading to more coherent and realistic videos. - **No Inference-Time Overhead**: RefAlign does not introduce any additional computational overhead during inference, making it highly efficient. ### Tech Stack: The project is implemented using Python, leveraging the PyTorch framework for training and inference. ### Why It Matters: RefAlign stands out as a significant advancement in the field of reference-to-video generation, offering practical solutions to common challenges in this domain. Its ability to improve reference fidelity, identity consistency, and text controllability makes it a valuable tool for researchers and practitioners in the field of video synthesis and generation. ### Repository: The project is hosted on GitHub, where it has garnered 351 stars, indicating its

Overview: MindPocket is an open-source, free, multi-platform, one-click deployable personal bookmark system with AI Agent integration. It simplifies the process of managing and finding saved content. Key features: - One-command deploy to Cloudflare Workers - Runs entirely on the Cloudflare free tier - Multi-platform support: Web, Mobile, and Browser Extension - AI Enhanced: RAG and AI Agent for smart tagging and summarization - CLI Ready: Official CLI for easy integration with external agents - Open Source: Fully open source, your data belongs to you Tech stack: TypeScript Why it matters: MindPocket streamlines the management of personal bookmarks, making it easy to find and manage saved content, while also integrating AI for enhanced tagging and summarization, all without the need for a server. Repository: jihe520/mindpocket GitHub stars: 352 GitHub description: Open-source, free, multi-platform, one-click deploy, AI Agent–integrated personal bookmarking system README excerpt: A fully open-source, free, multi-platform, one-click deployable personal bookmark system with AI Agent integration. MindPocket organizes your bookmarks with AI-powered RAG content summarization and automatic tag generation, making it easy to find and manage your saved content.

Overview: This monorepo provides LangChain and LangGraph components for various AWS services, aiming to replace and expand upon existing LangChain AWS components in the LangChain repository. Key features: - Includes LLM classes for AWS services like Bedrock and SageMaker Endpoints. - Supports vectorstores for Amazon MemoryDB, S3 Vectors, and AWS ElastiCache for Valkey. - Provides components for working with AWS Neptune graphs. - Includes Runnables for Amazon Bedrock Agents. - Offers custom checkpointing solutions for LangGraph agents using services like Bedrock AgentCore Memory and DynamoDB. - Enables secure code execution in isolated MicroVM environments for Amazon Bedrock AgentCore Code Interpreter. Tech stack: Python Why it matters: This repository aims to simplify and enhance the integration of LangChain with AWS services, making it easier and more efficient to build LangChain applications on AWS. Repository: langchain-ai/langchain-aws GitHub stars: 354

Overview: This repository is a comprehensive guide to building Retrieval-Augmented Generation (RAG) applications, offering a centralized directory of tools, libraries, and frameworks for RAG systems. Key features: - Courses and Learning Materials - Document Ingestor - Chunking Techniques - Retrieval - Query Transform - Agent Framework - Database - LLM - Embedding - Fine-tuning - LLM Observability - Prompt Techniques - Evaluation - User Interface - Complete RAG Applications Tech stack: Python, JavaScript, and various machine learning libraries. Why it matters: This guide provides a structured approach to building RAG applications, covering all necessary components and providing practical resources for each step of the process.

### Overview: ZigMa: A DiT-style Zigzag Mamba Diffusion Model This repository represents the official implementation of the paper titled "ZigMa: A DiT-style Zigzag Mamba Diffusion Model (ECCV 2024)". ### Key Features: - Zigzag scanning scheme for efficient parameter utilization. - Adaptation of the scanning scheme to video data. - Separation of spatial and temporal dimensions for improved parameter efficiency. - Incorporation of inductive bias for non-1D data. - Support for FP16 training and gradient checkpointing for better memory utilization and speed boosting. ### Tech Stack: Python ### Why it Matters: ZigMa offers a novel approach to diffusion models by combining the DiT-style Zigzag scanning scheme with Mamba diffusion, resulting in improved parameter efficiency and enhanced performance. This makes it particularly useful for handling non-1D data and achieving better results in diffusion models.

### Overview ZenCtrl is an all-in-one control framework designed for unified visual content creation using Generative Adversarial Networks (GANs). It generates high-resolution images from a single subject image without the need for fine-tuning, ensuring consistent subject preservation and control over shape, pose, and camera angle. ### Key Features - **No Fine-Tuning Needed**: Works directly from a single subject image. - **Control Over Shape, Pose, Camera Angle, Context**: Preserves the foreground fidelity while allowing precise control over the image elements. - **High-Resolution Multi-Scene Generation**: Supports multi-view and diverse-scene image creation. - **Modular Toolkit**: Provides preprocessing, control, editing, and post-processing functionalities. ### Tech Stack Python ### Why It Matters ZenCtrl is a significant advancement in image generation, offering a unified approach to visual content creation that can be seamlessly integrated into various workflows. Its ability to generate high-resolution images from a single subject image without the need for fine-tuning makes it highly versatile and practical for a wide range of applications, including product photography, fashion illustration, and more. ### GitHub Stars 351

Overview: Lale is a Python library designed for semi-automated data science, making it easy to automatically select algorithms and tune hyperparameters of pipelines compatible with scikit-learn. It adds value beyond scikit-learn by providing a consistent high-level interface, correctness checks, and interoperability with popular libraries. Key features: - Automatically selects algorithms and tunes hyperparameters of pipelines compatible with scikit-learn. - Consistent high-level interface for existing pipeline search tools. - Uses JSON Schema for correctness checks. - Growing library of transformers and estimators from popular libraries such as scikit-learn, XGBoost, PyTorch. Tech stack: Python Why it matters: Lale simplifies the process of semi-automated data science, making it accessible to data scientists who want to experiment with automated machine learning. It enhances the functionality of scikit-learn by providing automation, correctness checks, and improved interoperability, making it a valuable addition to the data science toolkit.

### Overview: This repository contains a collection of resources and links for mineral exploration and machine learning, focusing on practical applications and useful code examples. ### Key Features: - **Prospectivity**: Resources for mineral prospectivity analysis. - **Geology**: Data and tools related to geological exploration. - **Natural Language Processing (NLP)**: Techniques and resources for NLP in mineral exploration. - **Remote Sensing**: Tools and data for remote sensing applications in mineral exploration. - **Data Quality**: Resources for improving data quality in mineral exploration. - **Community**: Discussion forums and community resources. - **Cloud Providers**: Information on using cloud services for mineral exploration. - **Domains**: Resources for specific domains in mineral exploration. - **Overview**: General overview of the repository. - **Web Services**: Links to web services and APIs. - **Data Portals**: Access to various data portals related to mineral exploration. - **Tools**: Tools and libraries for machine learning in mineral exploration. - **Ontologies**: Ontologies and frameworks for machine learning in mineral exploration. - **Books**: Books on machine learning and mineral exploration. - **Datasets**: Datasets used in machine learning projects related to mineral exploration. - **Papers**: Papers and articles on machine learning in mineral exploration. - **Other**: Miscellaneous resources and links. ### Tech Stack: Python and various machine learning libraries such as TensorFlow, PyTorch, and scikit-learn. ### Why It Matters: These resources are essential for anyone involved in mineral exploration, providing practical tools and insights into machine learning applications in the

### Overview: This repository provides an overview of all components from the paper "Scaling Data-Constrained Language Models". Talks on the paper and slides used by speakers are available, along with experiments and models trained on varying data repetition and compute budgets. ### Key features: - Data: Repeating and Code - Models: Download, Training, Regular models, muP - Parametric Fit: Data - Downstream Evaluation: Rank Eval / Accuracy, Generative / Rouge, bAbI / Exact match - Plots & Tables: Plots, Tables ### Tech stack: Jupyter Notebook ### Why it matters: This repository provides a comprehensive overview of the scaling of data-constrained language models, including experiments, models, and data preparation, making it valuable for researchers and developers interested in the field of language models.

### Overview: DiffGAN-TTS is a PyTorch implementation of a Denoising Diffusion GAN-based text-to-speech (TTS) model. It aims to produce high-fidelity and efficient speech synthesis with minimal computational resources. ### Key Features: - **High-Fidelity Synthesis**: Utilizes Diffusion GANs for generating high-quality speech. - **Efficient Training**: Implements a shallow diffusion mechanism for faster training. - **Multi-Speaker Support**: Supports both single-speaker and multi-speaker TTS with external speaker embeddings. ### Tech Stack: - **Python**: The primary programming language used for the implementation. - **PyTorch**: Framework used for implementing the model and training process. ### Why It Matters: - **Efficiency**: Produces high-quality speech with minimal computational resources, making it suitable for real-time applications. - **Versatility**: Supports both single-speaker and multi-speaker TTS, catering to various use cases. ### Repository Status: - **Naive Version**: Available. - **Active Shallow Diffusion Mechanism**: Available for both **naive** and **shallow** models. ### Audio Samples: - Available at: [demo](https://github.com/keonlee9420/DiffGAN-TTS/tree/main/demo) ### Dependencies: - Install Python dependencies using `pip3 install -r requirements.txt`. ### Inference: - **Single-Speaker TTS**: ```bash python3 synthesize.py --text "YOUR_DESIRED_TEXT" --model MODEL --restore_step REST

Overview: MinusX is an AI Data Analyst that can be added to your Metabase, helping you ask and answer business questions in plain English. Key features: - Adds an AI side-chat to your Metabase - Uses your business context to reliably answer questions - Allows precise context specification - Supports custom metrics, dimensions, and sub-queries in Data catalogs Tech stack: TypeScript Why it matters: MinusX simplifies complex data analysis by providing intuitive answers to business questions, enhancing collaboration and decision-making in your organization.

Overview: PySS3 is a Python library designed for interpretable text classification using the SS3 model. It offers easy-to-use visualization tools and self-explanatory insights, making it a valuable tool for tasks requiring understanding model decisions. Key features: - SS3 class: Implements SS3 for text classification with a clean API. - Live_Test class: Interactively tests models in the browser with one line of code. - Classification support: Multi-label classification with `classify_multilabel()` method. - Visualization: Provides text fragments for understanding classification decisions. Tech stack: Python. Why it matters: PySS3 simplifies the process of working with SS3, making it accessible and effective for tasks requiring interpretability and explainability in text classification. Repository: sergioburdisso/pyss3 GitHub stars: 351 GitHub description: A Python library for Interpretable Machine Learning in Text Classification using the SS3 model, with easy-to-use visualization tools for Explainable AI.

### Overview: This project introduces DDPM-CD, a method that leverages Denoising Diffusion Probabilistic Models (DDPM) as feature extractors for remote sensing change detection. The repository includes pre-trained models and a lightweight change classifier that can be fine-tuned for various applications. ### Key Features: - Pre-trained DDPM model trained on off-the-shelf remote sensing images. - Generated images from the pre-trained model show common objects observed in remote sensing imagery. - Demonstrates the capability of diffusion models to extract essential semantics from training data. ### Tech Stack: Python ### Why it Matters: This project showcases the effectiveness of diffusion models as feature extractors, particularly in remote sensing change detection, providing a novel approach to change detection in remote sensing imagery.

### Overview: Construction Hazard Detection is an AI-driven project that enhances construction site safety by using YOLO for object detection. It identifies hazards such as workers without helmets or safety vests, and proximity to machinery or vehicles. The system clusters safety cone coordinates with HDBSCAN to create monitored zones and improves detection accuracy through post-processing algorithms. ### Key features: - **YOLO Object Detection**: Trained to detect workers without helmets or safety vests, machinery or vehicles too close to utility poles, and workers inside controlled areas. - **HDBSCAN Clustering**: Groups safety cone coordinates to create monitored zones. - **Post-Processing Algorithms**: Enhances detection accuracy through post-processing techniques. - **Real-Time Detection**: Detects hazards in real-time using live cameras. - **Supported Languages**: Traditional Chinese, Simplified Chinese, English, French, Thai, Vietnamese, Indonesian, and Japanese. ### Tech stack: - **Python 3.14** - **Ultralytics YOLO v8.4.115** - **HDBSCAN sklearn** - **FastAPI 0.128.0** - **Redis 7.1.0** - **Docker** - **Codecov** - **Roboflow Dataset Download** - **Hugging Face Model Repo** ### Why it matters: Construction sites often face significant risks due to workers' lack of protective gear or proximity to machinery. This project helps mitigate these risks by providing real-time alerts and improved safety monitoring, ensuring a safer working environment for all.

Overview: This project demonstrates a production-quality reference implementation for building Agentic applications that use Agents and Retrieval Augmented Generation (RAG) to query and interact with data stored in Google Cloud Databases. Key features: - A customer service agent (Cymbal Air) that manages flights and provides information about San Francisco International Airport. - RAG for reducing hallucinations by retrieving relevant data and augmenting prompts. - Agent-based orchestration for handling a wider variety of queries and breaking down complex questions. Tech stack: Python Why it matters: This project showcases the integration of Agents, RAG, and MCP Toolbox to create a robust and secure application for interacting with Google Cloud Databases, demonstrating best practices in AI and cloud database management.

Overview: This project aims to create a comprehensive application that leverages advanced AI models such as ChatGPT, TTS, STT, and other tools to facilitate language learning through interactive conversations and assessments. Key features: - Real-time language learning through oral conversations with ChatGPT - Speech recognition and assessment using advanced AI models - Integration with popular AI models for enhanced functionality - Support for various accents and languages - Speech synthesis capabilities with customizable voices Tech stack: Rust, Python, Web frameworks, AI models, and speech synthesis libraries. Why it matters: This project aims to provide a user-friendly and efficient language learning experience, utilizing cutting-edge AI technologies to enhance language learning and assessment.

Overview: SourceWeft is an open-source, self-hostable alternative to ChatGPT and NotebookLM designed to streamline AI workflows by connecting conversations, knowledge, working files, and task execution in one workstation. It allows users to bring their models, agents, skills, and infrastructure together, enabling multi-agent collaboration and sandboxed execution. Key features: - Multi-agent collaboration and sub-agents - Skill library for building on skills - MCP & connected tools - Sandboxed execution - Multiple models & BYOK - Across devices & remote access - Creation & working files - Knowledge & verifiable citations - Self-hosting & team collaboration Tech stack: TypeScript Why it matters: SourceWeft provides a flexible, open-source solution for AI workstations, allowing users to integrate their own models, skills, and infrastructure, and collaborate across devices and teams. Repository: SourceWeft/SourceWeft GitHub stars: 330 GitHub description: Open-source, self-hostable alternative to ChatGPT and NotebookLM — agents, skills, MCP, and cited knowledge in one AI workstation. Your models, your infrastructure. Free to use GPT 6: https://SourceWeft.com/

Overview: This repository contains the official PyTorch implementation of TreeLoRA, an efficient continual learning method for Large Language Models (LLMs) that uses layer-wise LoRA adapters guided by a hierarchical gradient-similarity tree. The repository includes code for various model implementations, training scripts, and utility functions, making it a comprehensive resource for researchers and developers working on continual learning and large language models. Key features: - Efficient continual learning through layer-wise LoRA adapters - Hierarchical gradient-similarity tree for adapter organization - Support for multiple LLM architectures (Gemma, LLaMA, Mistral, etc.) - DeepSpeed integration for efficient training - Flash attention implementation for improved performance Tech stack: Python Why it matters: This implementation is significant for researchers and developers working on continual learning and large language models, as it provides a practical and efficient method for updating and learning from new data in large language models. Repository: ZinYY/TreeLoRA GitHub stars: 350 GitHub description: [ICML 2025] A pytorch implementation of the paper "TreeLoRA: Efficient Continual Learning via Layer-Wise LoRAs Guided by a Hierarchical Gradient-Similarity Tree".