Projects

Overview: sdwebui-java-sdk is a Java library that provides a convenient interface for integrating Stable Diffusion Webui into Java applications, making it easier to generate images and text-to-image functionalities. Key features: - Integrates with Stable Diffusion Webui for seamless image generation. - Supports text-to-image and image-to-image functionalities. - Simplifies the process of creating and using the SD Webui SDK. Tech stack: Java Why it matters: This library streamlines the process of integrating Stable Diffusion Webui into Java applications, allowing developers to easily generate images and perform text-to-image functionalities without needing to manually interact with the Webui.

Overview: WizSearch is an intelligent tool designed for effortless data retrieval across documents and seamless browsing the web. Key features: - Seamless browsing across documents and the web. - Uses advanced LLM/VLM models for natural language understanding and generation. - Supports open-source and closed-source models for natural language processing. - Enhances search relevance using embeddings and vector databases. - Provides advanced search capabilities with intelligent search. - Converts documents to text using MarkItDown. - Supports efficient data storage and retrieval with Qdrant. - Monitors and observability with Langfuse. - Uses Streamlit for an interactive and user-friendly interface. Tech stack: Python. Why it matters: WizSearch simplifies the process of finding information efficiently and seamlessly across various documents and online resources, making it a valuable tool for developers and professionals alike.

Overview: This project is an Insurance AI Assistant system that integrates PostgreSQL, Milvus, and specialized AI agents for Life, Home, and Auto insurance domains. It uses real-time database synchronization, semantic search via OpenAI embeddings, and a Streamlit UI for user interaction. Key features: - Multi-Database Integration: PostgreSQL for structured data and Milvus for vector embeddings - Real-time Sync: Continuous database synchronization between PostgreSQL and Milvus - Specialized AI Agents: Domain-specific insurance experts for Life, Home, and Auto - Intelligent Routing: Smart query routing to appropriate specialists - Streamlit UI: Beautiful, interactive web interface Tech stack: Python Why it matters: This system is designed for insurance tech demos or customer service augmentation, providing a sophisticated AI-powered solution for insurance queries. Repository: SamJoeSilvano/Insurance-Agentic-AI GitHub stars: 34 Primary language: Python GitHub description: A sophisticated AI-powered insurance assistant system that combines PostgreSQL database management, vector embeddings with Milvus, and specialized AI agents for insurance domain expertise.

### Overview: This project is a minimal implementation of text diffusion in Python, inspired by Karpathy's MicroGPT. It aims to provide a clear and concise explanation of the text-diffusion loop without relying on large frameworks or datasets. The project is designed for readers who want to understand the core concept without the complexity of production systems. ### Key Features: - **Minimal Implementation**: Pure Python implementation with a 150-line script. - **Mask Tokens**: Instead of shifting tokens, it uses masking to propagate noise. - **Full Sequence Denoising**: Denoises the entire sequence in one go, not just predicting left to right. - **Bidirectional Attention**: Unlike autoregressive models, it uses a bidirectional approach to predict the full sequence. - **Smaller Scripts**: Smaller versions (`train_minimal.py`, `train_pure.py`) are script-safe and CLI-driven. ### Tech Stack: - **Python**: The primary language used for development. - **NumPy**: Used for numerical operations and script safety. - **PyTorch**: Optional for a bidirectional transformer version. ### Why It Matters: This project aims to demystify the text-diffusion process by providing a clear, concise explanation of the core concept. It is intended for readers who want to understand the underlying mechanics of text generation without the complexity of production systems. ### Repository: - **Primary Language**: Python - **GitHub Stars**: 35 - **Repository**: Siwoo4985/Micro-Diffusion - **GitHub Description**: The essence

Overview: This repository contains resources and updates related to our survey paper "Efficient Diffusion Models: A Comprehensive Survey from Principles to Practices", published in TPAMI 2025. Key features: - Survey paper on efficient diffusion models - Open-source and closed-source reproduction projects - Timeline of efficient diffusion models - Conceptual framework of efficient diffusion models - Comparison of core features of mainstream diffusion-based generative models - Various adapters and their applications Tech stack: Python Why it matters: The repository provides valuable resources for researchers and practitioners interested in understanding and implementing efficient diffusion models. GitHub stars: 35 Repository: TsinghuaC3I/Efficient-Diffusion-Models

Overview: SteinDreamer is a project that introduces variance reduction techniques for text-to-3D score distillation using the Stein identity, aiming to improve the accuracy and efficiency of 3D score distillation. Key features: - Variance reduction for text-to-3D score distillation - Utilizes the Stein identity for effective variance reduction - Improves the accuracy and efficiency of 3D score distillation Tech stack: Python, PyTorch, and TensorFlow Why it matters: This project offers a novel approach to text-to-3D score distillation, potentially leading to more accurate and efficient 3D score generation, which could have significant applications in computer vision and related fields.

Overview: This project is a PyTorch implementation of a Zero-Shot Structure-Preserving Diffusion Model for High Dynamic Range Tone Mapping, presented at CVPR 2024. Key features: - Zero-Shot Structure-Preserving Diffusion Model - Training and testing on Flickr2K dataset - Pre-trained models available for download Tech stack: Python Why it matters: This project offers a flexible and effective approach to tone mapping using structure-preserving diffusion models, showcasing advancements in high dynamic range tone mapping techniques.

### Overview: ARCQuant is a high-performance quantization framework designed specifically for low-bit low-precision language models (LLMs). It addresses the challenge of accuracy degradation in critical channels during fine-grained quantization formats like NVFP4, by leveraging augmented residual channels (ARC) to enhance quantization performance without compromising the inference pipeline. ### Key Features: - **ARCQuant**: A framework that boosts NVFP4 quantization accuracy by incorporating augmented residual channels. - **Fine-Grained Quantization**: NVFP4 isolates quantization noise, but residual channels can cause severe accuracy degradation. - **Augmented Residual Channels (ARC)**: Capture quantization residuals of critical channels and fuse them back into the computation as augmented residual channels, thereby improving accuracy. ### Tech Stack: - **Primary Language**: Cuda - **GitHub Stars**: 35 ### Why It Matters: ARCQuant is crucial for maintaining high accuracy in low-bit quantized models, ensuring that LLMs perform well across a wide range of applications without sacrificing performance. By addressing the specific challenges of fine-grained quantization, it provides a robust solution for quantizing LLMs effectively. ### Repository: - **Repository**: actypedef/ARCQuant - **Primary Language**: Cuda - **GitHub Stars**: 35 - **GitHub Description**: [ACL 2026 Main] Code for the paper "ARCQuant: Boosting NVFP4 Quantization with Augmented Residual Channels for LLMs" ### Installation: - **Requirements**: Ensure CUDA 12.

### Overview: This project is an implementation of the paper "2D3MF: Deepfake Detection using Multi Modal Middle Fusion". The code and models are available for deepfake detection using a multi-modal middle fusion approach. ### Key Features: - **2D3MF Model**: Implements a deepfake detection model using a multi-modal middle fusion technique. - **PyTorch**: The primary framework used for the implementation. - **GitHub Stars**: 35, indicating the project has gained some traction in the community. ### Tech Stack: - Python 3.7 - 3.12 - PyTorch 1.11 - Torchvision 0.12 ### Why It Matters: This project provides a practical solution for detecting deepfakes, which are manipulated videos that can be difficult to distinguish from real videos. The multi-modal middle fusion approach enhances the model's ability to recognize subtle differences between real and fake videos, making it a valuable tool in the fight against deepfake content. ### Repository Structure: The repository is organized into several directories, including: - `assets`: Contains images for the README. - `LICENSE`: The license for the project. - `README.md`: The main documentation. - `MODEL_ZOO.md`: A placeholder for model zoo information. - `.gitignore`: A file to ignore certain files and directories. ### Installation and Running: To use the model, you need to install PyTorch and Torchvision. The project includes a `requirements.txt` file that lists the necessary dependencies. You can install the model using pip:

### Overview Cora is a novel image editing method that achieves high-quality results in just four diffusion steps. Unlike traditional fast methods, Cora uses semantic correspondences between the original and edited images to preserve structure and appearance. It is fast, controllable, and requires no additional training, making it an ideal choice for researchers and practitioners. ### Key Features - **Few-Step Editing**: Delivers high-quality results in just 4 diffusion steps, making it fast and memory-efficient. - **Structure-Aware**: *Correspondence-aware Latent Correction* (CLC) aligns noise terms to the target layout, eliminating ghosting artifacts. - **Appearance Control**: *Attention Interpolation* lets you blend source appearance with the prompt via a controllable parameter (α). - **Structure Control**: *Query Matching* maintains or relaxes spatial layout via a controllable parameter (β). ### Tech Stack - **Python 3.10+** - **PyTorch 2.5.1** ### Why It Matters - **Few-Step Editing**: Delivers high-quality results in just 4 diffusion steps, making it fast and memory-efficient. - **Structure-Aware**: *Correspondence-aware Latent Correction* (CLC) aligns noise terms to the target layout, eliminating ghosting artifacts. - **Appearance Control**: *Attention Interpolation* lets you blend source appearance with the prompt via a controllable parameter (α). - **Structure Control**: *Query Matching* maintains or relaxes spatial layout via a controllable parameter (β). ###

Overview: This project is a Streamlit application that integrates Amazon Bedrock Knowledge Bases with multimodal Large Language Models (LLMs), allowing users to ask questions and receive relevant responses based on the selected knowledge base and LLM. Key features: - Amazon Bedrock Knowledge Bases integration - Multimodal LLM support - Image upload capability - AWS Bedrock model access configuration - Customizable knowledge base creation and deletion Tech stack: Python Why it matters: This application leverages the power of Bedrock Knowledge Bases and multimodal LLMs to provide a more comprehensive and contextually relevant response to user queries, enhancing the user experience in a conversational and interactive manner.

### Overview: This repository contains a survey paper titled "Towards Rationality in Language and Multimodal Agents: A Survey" presented at the NAACL 2025 conference. The paper explores the concept of rationality in language and multimodal agents, emphasizing the importance of consistent and reliable decision-making in critical sectors like healthcare and finance. ### Key Features: - **Define Rationality**: A comprehensive overview of rationality in the context of language and multimodal agents. - **Towards Rationality in Agents**: A survey of the notion of rationality in language and multimodal agents. - **Evaluating Rationality in Agents**: Methods and techniques for evaluating rationality in agents. ### Tech Stack: Python, Markdown, and LaTeX. ### Why it Matters: Rationality is crucial for agents in critical sectors like healthcare and finance, where consistent and reliable decision-making is essential. This survey provides a foundational understanding of rationality, which is essential for the development of agents that can make informed decisions based on evidence and logical principles. Repository: bowen-upenn/Agent_Rationality GitHub Stars: 35 GitHub Description: [NAACL 2025] Towards Rationality in Language and Multimodal Agents: A Survey

### Overview: FedLLM Factory is a unified library for LoRA-based federated LLM fine-tuning, supporting 10+ baselines and enabling efficient and privacy-preserving fine-tuning of large language models across heterogeneous clients. ### Key Features: - **10+ Baselines**: FedIT, FFA-LoRA, FLoRA, FlexLoRA, FedDPA, FedSA-LoRA, FedEX-LoRA, FedSVD, RAVAN, RoLoRA, FedLEASE, SLoRA, ILoRA, FedRotLoRA, HiLoRA, FedMomentum. - **Federated Fine-tuning**: Supports asynchronous federated learning for fine-tuning large language models. - **Easy Dataset Generation**: Provides a simple frontend for dataset generation, including NLU (SST-2, IMDB), QA (Dolly), and Mathematical Reasoning (GSM8K). - **Configuration Flexibility**: Configurable through `config.yaml` and `utils/options.py`, allowing for easy adjustment of parameters. ### Tech Stack: Python ### Why It Matters: FedLLM Factory simplifies the process of fine-tuning large language models in a federated learning environment, making it easier to experiment with various baselines and achieve efficient privacy-preserving fine-tuning across diverse client environments.

Overview: `CleverBird` is a Swift package designed for seamless integration with OpenAI's API, providing advanced chat capabilities and token management in applications. Key features: - Asynchronous API calls with Swift's async/await syntax - Streamed responses for real-time generated content - Built-in token counting for usage limit calculations - Token Encoding: Facilitates token counting and encoding through the `TokenEncoder` class - Document Embedding and Similarity Queries: Utilize the `EmbeddedDocumentStore` class for managing and querying document similarities Tech stack: Swift Why it matters: `CleverBird` simplifies the integration of OpenAI's chat APIs, making chat interactions more efficient and user-friendly, while also offering enhanced functionality through document embedding and similarity queries.

### Overview: VOXRAD is a voice transcription application designed for radiologists, utilizing locally deployed ASR and large language models to restructure and format reports as per predefined user instruction templates. ### Key Features: - Voice transcription - Report formatting - Integration with large language models - Customizable templates - Potential for dictating other structured notes ### Tech Stack: Python ### Why It Matters: VOXRAD offers a streamlined and efficient way for radiologists to transcribe and format reports, enhancing accuracy and streamlining the reporting process.

Overview: This repository contains code and information related to fine-tuning lightweight language models for the text classification task, focusing on Phi-3 and H2O-Danube models. Key features: Fine-tuning of lightweight language models for text classification, using Phi-3 (3.8B) and H2O-Danube (1.8B) models. Tech stack: Jupyter Notebook. Why it matters: These models are efficient, fast, and perform well, making them suitable for text classification tasks without requiring significant computational resources.

### Overview: Open Telco (OTel) AI is an open ecosystem of telecom datasets, benchmarks, models, and reproducible training and inference recipes. The project aims to provide a shared, open-source AI foundation that allows telecom operators to choose their models, software, hardware, and infrastructure. ### Key features: - Multiple training recipes spanning LLM post-training, embedding, reranking, and classification. - Multiple inference stacks covering LLM generation, embeddings, and reranking, with workload support varying by stack. - Examples of SFT, CPT, DPO, GRPO, PPO, OSFT, LoRA, QLoRA, and DoRA. - Documentation, dependency requirements, runnable entry points, and sample workflows. - Support for multiple software stacks across hardware from AMD, NVIDIA, AWS, Apple, Tenstorrent, and Intel. - Telecom datasets, model releases, and reported evaluation results documented in the repository and accompanying Hugging Face model and dataset cards. ### Tech stack: Python ### Why it matters: OTel AI provides a shared, open-source AI foundation that empowers telecom operators to choose their models, software, hardware, and infrastructure, giving them the freedom to adapt and deploy AI on their own terms.

### Overview: TableDiffusion is a project focused on providing differentially-private generative models for sensitive tabular data. The goal is to enable the synthesis of data that maintains the statistical properties of the original dataset while ensuring the privacy of individuals' information. ### Key Features: - `TableDiffusion`: The first differentially-private diffusion model for tabular data. - `TableDiffusion` is a project focused on generating differentially-private tabular datasets. ### Tech Stack: Python ### Why It Matters: This project is significant for researchers and developers looking to create privacy-preserving datasets. It provides a practical solution for generating synthetic data while ensuring that individual privacy is maintained, which is crucial in fields such as data science, machine learning, and research.

### Overview: ParallelHue is a truthful scheduler-step and streaming visualization tool for local LLM inference, designed to provide a clear and accurate representation of token provenance and streaming behavior. ### Key Features: - **Scheduler and Detokenizer Telemetry**: Tracks and visualizes token provenance and role for each emitted token. - **Exact Color Mode**: Displays verified generation steps with exact colors, requiring token provenance, detokenizer traces, and token IDs. - **SSE Chunk Mode**: Shows ordinary streaming chunks with monochrome colors, without verified speculative-step telemetry. - **Four-Color Cycle**: Tracks step progression in a shared four-color cycle, with each step retaining its color even when token roles differ. - **Monochrome for Non-Verifiable Backends**: Uses monochrome colors for non-verified speculative decoding backends (`generic` and `auto`). - **Semantic Boundary**: Keeps a clear semantic boundary between exact and SSE chunk modes. ### Tech Stack: Python ### Why It Matters: - **Transparent Token Provenance**: Provides a clear and accurate representation of token provenance and role, enhancing transparency and accountability. - **Detailed Streaming Behavior**: Shows the detailed streaming behavior of each request, including concurrent contribution, stalls, and speculative bursts. - **Multi-Agent Workload**: Ideal for multi-agent, parallel, and batched workload environments, where aggregate throughput is a key performance metric. ### Repository: hikarioyama/ParallelHue ### GitHub Stars: 35 ### Primary Language: Python ### Installation: ```sh python

Overview: QUAYNOR is a lightweight, blazing fast AI inference engine written in Rust, designed for mobile devices. It supports offline inference, streaming completions, and is compatible with various programming languages including Python, Flutter, React Native, Swift, and Kotlin/Android. Key features: - Chat with streaming completions - Minijinja chat templates - Tokenizer helpers - Optional embeddings and cross-encoder reranking - Grammar-based tool calling Tech stack: Rust Why it matters: QUAYNOR provides a production-oriented solution for AI inference, allowing developers to integrate AI capabilities into their applications without relying on external inference APIs or cloud storage. It simplifies the process of using AI models on mobile devices, ensuring that models stay on disk or can be loaded from trusted sources, and supports streaming responses and bounded context sizing. GitHub stars: 35 Repository: iBz-04/quaynor