Projects

Overview: LiteGraph is a multi-modal AI data platform that combines property graph, relational, and vector support, designed to power knowledge and AI persistence and retrieval. Key features: - Property graph database with relational, vector, and MCP support - Embedded .NET graph library targeting .NET 8.0 and 10.0 - SQLite and PostgreSQL providers - REST server, MCP server, and console tools - Graph algorithms and transactions - Vector search on PostgreSQL and SQLite - Official C#, Python, and JavaScript SDKs - Next.js/React dashboard - Docker Compose deployments for single-node SQLite, PostgreSQL, and multi-node MCP clusters Tech stack: C# Why it matters: LiteGraph simplifies AI data management by integrating multiple storage and retrieval methods, making it a versatile solution for applications requiring graph relationships, tags, labels, JSON data, and vector search. Repository: litegraphdb/litegraph GitHub stars: 132

Overview: 本书提供了Amazon Bedrock生成AI应用程序开发入门的样例代码，适合书中的动手实践。这本书是《Amazon Bedrock 生成AI 应用程序开发入门》的配套资源，旨在帮助读者更好地理解和掌握相关技术。 Key features: - 本书包含章节级别的样例代码，以及一些不易手动修改的配置参数。 - 曾经购买过该书的读者可以轻松地复制和粘贴代码。 - 代码中列出了未来可能存在的功能更新和错误修正的说明。 - 曾经使用过AWS Cloud9的读者需要了解如何使用AWS Cloud9的新版本。 Tech stack: Python Why it matters: 本书是《Amazon Bedrock 生成AI 应用程序开发入门》的配套资源，旨在帮助读者更好地理解和掌握相关技术。通过本书的样例代码，读者可以更直观地学习和实践生成AI应用程序开发。 Repository: minorun365/bedrock-book Primary language: Python GitHub stars: 132 GitHub description: 書籍「Amazon Bedrock 生成AIアプリ開発入門」のサンプルコード README excerpt: # 書籍「Amazon Bedrock 生成AIアプリ開発入門」 ハンズオン用サンプルコード 表題の書籍のハンズオンを実施しやすいよう、サンプルコード部分をファイルとして格納したリポジトリです。 - 紙の書籍を購入くださった方も

### Overview: oc8 is an open-source, self-hosted AI agent orchestration platform designed to run governed AI agents across various business systems, including ERP, CRM, Microsoft 365, and others. It ensures secure collaboration and compliance, with isolated AI agents, a knowledge base management system, and human approval controls. ### Key Features: - 🔒 Isolated AI Agents: Each agent runs in its own container, with access only to what oc8 explicitly grants it. - 📚 RAG Knowledge Base Management: Retrieval-augmented knowledge, scoped per agent or department. - 🧠 Works With Any LLM: Anthropic, OpenAI, OpenRouter, local Ollama, or any OpenAI-compatible endpoint. - 🛡️ Sovereign by Design: Self-hosted; your data and credentials never leave infrastructure you control. - 🏢 Scoped and Built for Enterprise Usage: Tenant isolation, audit trails, and approval thresholds from day one. - 🔌 Works With Your Existing Stack: oc8 orchestrates the CRM, ERP, and tools you already run; nothing to rip out and replace. - 🧑‍⚖️ Human in Control: Every consequential action waits for approval unless you've explicitly said otherwise. - 🧩 Extensible in Every Direction: The capa system: connectors, skills, tool packs, and department templates as installable folders. - 🔗 Built on Open Protocols: Every tool call goes through the same governed MCP gateway, compatible with standard agent tooling. - 📈 Built to Scale:

Overview: AnyControl is a multi-control image synthesis model that allows users to input various control signals to generate natural and harmonious results. Key features: - Supports any combination of user-provided control signals. - Novel Multi-Control Encoder architecture for comprehensive understanding of complex multi-modal inputs. - High-quality and high-fidelity results under versatile control signals. Tech stack: Python Why it matters: AnyControl enables artists and designers to create unique and dynamic artwork by incorporating a wide range of control signals, resulting in visually appealing and contextually relevant images.

### Overview: Angle is a Python implementation of the English language, designed to facilitate programming computers through voice. It compiles to Python bytecode and is fully debuggable, making it a drop-in-replacement for classic Python. The language is currently in development and can be run directly in WASM via wasp. ### Key features: - **Natural Syntax**: Angle's syntax is designed to mimic natural language, making it accessible to many more programmers. - **Brevity and Readability**: The language emphasizes brevity and readability, with verbosity optional. - **Type and Sigils Optional**: Types and sigils are optional, reducing verbosity. - **Voice Programming**: The primary purpose is to facilitate programming via voice, making it accessible to non-technical users. ### Tech stack: Python ### Why it matters: Angle is the first speakable programming language, making it accessible to many more people. It aims to simplify software development by reducing verbosity and making the language more natural and readable. ### Tech Stack: - **Python**: The primary language used for development and implementation. - **WASP**: A Rust implementation that compiles to WebAssembly, enabling direct execution in browsers and other environments. ### Installation: - **Using pip**: `pip install angle` - **From Source**: `git clone --recursive git@github.com:pannous/angle.git` and `./install.sh` ### Examples: - **Basic Operations**: ```python assert two minus 1.5 equals 0.5 beep three times

Overview: This repository is a collection of vibrant multimodal research, focusing on advancements propelled by large language models (LLMs). It covers domains such as Vision, Audio, Agent, Robotics, Fundamental Sciences, and Ominous. Key features: - Image - Video - 3D - Document - Vision Encoder Tech stack: Unknown Why it matters: This repository provides a curated list of works and resources related to large language models in the field of multimodal research, making it a valuable resource for researchers and practitioners interested in the latest advancements in this area.

Overview: MirrorGAN is a Pytorch implementation of the MirrorGAN paper, focusing on text-to-image generation through redescription. It includes pretrained modules for STEM and STREAM, and provides scripts for both training and testing the model. Key features: - Pretrained STEM and STREAM modules - Text-to-image generation via redescription - Pytorch implementation Tech stack: Python Why it matters: MirrorGAN offers a practical approach to text-to-image generation, demonstrating the potential of redescription techniques in generating high-quality images from textual descriptions. Repository: qiaott/MirrorGAN Primary language: Python GitHub stars: 132 GitHub description: Pytorch implementation of MirrorGAN

### Overview: LitePali is a lightweight, efficient implementation of ColPali specifically designed for image-based document retrieval and indexing. Optimized for cloud deployment, it aims to provide fast and accurate search capabilities using state-of-the-art vision-language models. ### Key Features: - **Minimal Dependencies**: Uses minimal system requirements, making it easy to deploy. - **Direct Image Processing**: No need for complex PDF parsing, focusing solely on image processing. - **Deterministic File Processing**: Ensures consistent results through deterministic file processing. - **Batch Processing**: Efficiently processes multiple files in a batch for improved performance. ### Tech Stack: - **Python**: The primary language used for development. - **Pytorch**: Version 1.8+ for improved performance. ### Why It Matters: LitePali is designed to efficiently retrieve and index document images, making it ideal for applications requiring rapid and accurate document search capabilities. Its focus on image processing and optimized cloud deployment make it a valuable tool in the field of document retrieval and indexing. ### GitHub Stars: 132 ### Repository: - Repository: s-emanuilov/litepali - Primary Language: Python - GitHub Stars: 132 ### Installation: ```bash pip install litepali ``` ### Usage: ```python from litepali import LitePali, ImageFile # Initialize LitePali litepali = LitePali() # Add some images with metadata and page information litepali.add(ImageFile( path="path/to/image1.jpg", document_id

### Overview: WorldMM is a dynamic multimodal memory agent designed for long video reasoning. It constructs multimodal, multi-scale memories that capture both textual and visual information, and employs adaptive retrieval across multiple memories with reasoning. ### Key Features: - **Dynamic Memory Construction**: WorldMM dynamically builds three memory modules—episodic, semantic, and visual—to support long-term reasoning. - **Adaptive Retrieval**: It employs adaptive retrieval across multiple memories to enhance reasoning capabilities. - **Versatile Backbones**: Supports various backbone models for retrieval and reasoning, including `gpt-5` and `qwen3vl-8b`. ### Tech Stack: - Python ### Why It Matters: WorldMM offers a novel approach to dynamic multimodal memory agents, making it particularly useful for long video reasoning tasks. Its ability to capture both textual and visual information and adaptively retrieve across multiple memories makes it highly effective for complex reasoning tasks.

Overview: DiffuseKronA is a parameter-efficient fine-tuning method designed for personalized diffusion models, achieving up to a 75% reduction in parameters compared to LoRA-DreamBooth while maintaining comparable results in stability and text alignment fidelity. Key features: - Parameter-efficient: A minimum 35% reduction in parameters, with up to 75% reduction possible. - Enhanced stability: More stable compared to LoRA-DreamBooth, making it easier to fine-tune. - Improved text alignment and fidelity: Captures better subject semantics and large contextual prompts. - Enhanced interpretability: Uses the Kronecker product to capture structured relationships in attention-weight matrices, making it more interpretable. Tech stack: Python Why it matters: DiffuseKronA offers significant improvements in parameter efficiency, stability, and interpretability, making it a valuable tool for fine-tuning diffusion models tailored to specific applications.

Overview: This project is the official repository for "Bad Actor, Good Advisor: Exploring the Role of Large Language Models in Fake News Detection," a paper accepted by AAAI 2024. The dataset and code are available in the `data` and `code` folders, respectively. Key features: - Pretrained models for Chinese and English datasets. - Requirements for Python 3.10.13 and CUDA 11.3. - Scripts for running the models on the Chinese and English datasets. Tech stack: Python with CUDA 11.3. Why it matters: This project provides a comprehensive solution for detecting fake news using large language models, making it a valuable resource for researchers and developers in the field of AI and misinformation detection.

### Overview: UMbreLLa is a cutting-edge project that leverages speculative decoding and quantization techniques to deploy large language models (LLMs) efficiently on consumer devices. This project aims to achieve performance comparable to human reading speed, particularly excelling in coding tasks. ### Key Features: - **Speculative Decoding**: Predicts the next token in a sequence to improve inference speed. - **Quantization**: Reduces the model's memory footprint and computational requirements. - **Offloading**: Moves the model's inference to a dedicated hardware accelerator to enhance performance. ### Tech Stack: - Python ### Why It Matters: UMbreLLa significantly improves the efficiency and responsiveness of large language models, making them suitable for real-world applications such as coding tasks. By leveraging speculative decoding and quantization, it achieves performance comparable to human reading speed, making it a valuable tool for developers and engineers. ### Repository: - **GitHub Repository**: Infini-AI-Lab/UMbreLLa - **Stars**: 131 - **Description**: LLM Inference on consumer devices ### Primary Language: - Python ### GitHub Description: UMbreLLa combines offloading, speculative decoding, and quantization to efficiently deploy large language models on consumer devices. It achieves performance comparable to human reading speed, particularly excelling in coding tasks. ### Benchmarks: - **Throughput**: - RTX 4090: Llama3.1-70B-Instruct-AWQ: 7.2 tokens/sec, Llama3.1-

### Overview DSBench is a benchmark designed to evaluate data science agents by simulating real-world data analysis and modeling tasks. It provides a comprehensive set of tasks and data files that a data science agent must solve, ensuring that the agents' capabilities are tested under realistic conditions. ### Key Features - **Realistic Tasks**: Tasks are sourced from modeloff and Kaggle, including both data analysis and data modeling. - **Task Instructions**: Agents are given clear task instructions that may include images and tables. - **Solution Generation**: Agents must generate a solution that addresses the task, resolving the described problem. ### Tech Stack - **Programming Languages**: Jupyter Notebook - **Data Sources**: Modeloff, Kaggle ### Why It Matters DSBench serves as a crucial tool for researchers and developers in the field of data science, providing a standardized way to evaluate the capabilities of data science agents. By simulating real-world scenarios, it helps in identifying the strengths and weaknesses of different agents, ultimately contributing to the advancement of data science expertise. Repository: LiqiangJing/DSBench GitHub Stars: 130 GitHub Description: [ICLR 2025] DSBench: How Far are Data Science Agents from Becoming Data Science Experts?

### Overview: The LMT project aims to advance the frontier of Multilingual Machine Translation (MMT) by building **Inclusive**, **Scalable**, and **High-performance** multilingual translation models. ### Key features: - **Inclusive Models**: Covering 60 languages and 234 translation directions. - **Scalable Models**: Available in four sizes: 0.6B / 1.7B / 4B / 8B. - **High-performance Models**: Built with large-scale multilingual translation models. ### Tech stack: Python ### Why it matters: The LMT project introduces innovative solutions to address the challenges of deploying large multilingual models, making them more accessible and efficient for a wider range of applications.

Overview: JevAny is an open infrastructure designed for training, evaluating, and deploying System 1 decision models across various languages and multimodal backbones. It supports the adaptation and evaluation of models using released checkpoints or training on custom data. Key features: - Model adaptation and evaluation - Training on custom data - Integration with different language and multimodal backbones - Support for a wide range of applications Tech stack: Python Why it matters: JevAny simplifies the process of training and deploying decision models, making it accessible for a variety of use cases, including integrating with different language and multimodal backbones, and providing a platform for evaluating and adapting models on custom data. Repository: SimpleJev/JevAny GitHub stars: 126 GitHub description: Open infrastructure for training, evaluating, and deploying System 1 decision models across language and multimodal backbones.

### Overview: Dataset Diffusion is a project that introduces a novel method for generating pixel-level semantic segmentation labels using the text-to-image generative model Stable Diffusion (SD). By appending class prompts, utilizing cross-attention, and applying self-attention exponentiation, the project aims to create segmentation maps that correspond to synthetic images, serving as pseudo-labels for training semantic segmenters. This approach significantly reduces the need for labor-intensive pixel-wise annotation, making it particularly useful for preparing training data for deep vision models. ### Key Features: - **Class-Prompt Appending**: Enhances the ability to generate segmentation maps corresponding to synthetic images. - **Class-Prompt Cross-Attention**: Utilizes cross-attention to improve the quality of generated segmentation maps. - **Self-Attention Exponentiation**: Applies self-attention exponentiation to further refine the generated segmentation maps. ### Tech Stack: - **Python 3.8+** - **PyTorch ≥ 1.10** - **HuggingFace installations**: diffusers, transformers, safetensors ### Why It Matters: - **Efficiency**: Reduces the need for labor-intensive pixel-wise annotation, making it ideal for preparing training data for deep vision models. - **Performance**: Utilizes the text-to-image generative model Stable Diffusion to generate high-quality segmentation maps, serving as pseudo-labels for training semantic segmenters. - **Versatility**: Works across multiple datasets, including PASCAL VOC and MSCOCO, demonstrating broad applicability. ### Repository: - **GitHub Repository**: VinAIResearch/Dataset

### Overview: Genome-Factory is an integrated Python library designed for tuning, deploying, and interpreting genomic models. It supports a variety of models and preprocessing techniques, making it a versatile tool for genomic research and analysis. ### Key features: - **Genome Collector**: Acquires genomic sequences from public repositories. - **Model Loader**: Supports major genomic models and their tokenizers. - **Model Trainer**: Configures workflows, adapts models, and executes training with fine-tuning or parameter-efficient methods. - **Inference Engine**: Enables embedding extraction and sequence generation. - **Benchmarker**: Provides standard benchmarks and allows integration of custom evaluation tasks. - **Biological Interpreter**: Enhances interpretability through sparse auto-encoders. ### Tech stack: Python ### Why it matters: Genome-Factory simplifies the process of genomic model tuning, deployment, and interpretation, making it an essential tool for researchers and developers in the field of genomics.

Overview: This repository contains the official code and instructions for the paper "Revisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark". The project aims to provide insights into the capabilities, limitations, and principles of zeroth-order optimization methods for reducing memory costs during large language model fine-tuning. Key features: - The project includes the source code and guide for a benchmark study on zeroth-order optimization methods. - It covers five different language model families (Robert, OPT, LLaMA, Vicuna, and Mistral). - It includes three different levels of task complexity (binary classification, question-answering, and commonsense reasoning). - It provides four fine-tuning schemes (full fine-tuning, LoRA, prefix tuning, and prompt tuning). - It includes six zeroth-order optimization methods (ZO-SGD, ZO-SGD-Sign, ZO-SGD-MMT, ZO-SGD-Cons, ZO-Adam, and forward gradient). - It also introduces three novel enhancements to zeroth-order optimization (block-wise descent, hybrid training, and gradient sparsity). Tech stack: Python Why it matters: This project offers researchers a comprehensive resource to better understand and evaluate the performance of zeroth-order optimization methods in the context of memory-efficient LLM fine-tuning, thereby contributing to the field's development and practical applications.

Overview: This project integrates Automatic Speech Recognition (ASR) functionality into Unity games using the Vosk library, enabling seamless integration with speech-to-text capabilities. Key features: - Real-time speech recognition - Unity integration - Vosk library support Tech stack: C# Why it matters: This project simplifies the process of adding speech recognition capabilities to Unity games, making it easier for developers to incorporate voice-based interactions into their projects.

Overview: DMax is a new paradigm for efficient dLLMs that mitigates error accumulation in parallel decoding, enabling aggressive decoding parallelism while preserving generation quality. It reformulates decoding as a progressive self-refinement from mask embeddings to token embeddings, and introduces On-Policy Uniform Training and Soft Parallel Decoding to enhance performance. Key features: - Aggressive Parallel Decoding: Achieves 6.0 TPF on math and reasoning tasks and 6.6 TPF on code tasks while preserving accuracy. - Self-Revising dLLM: Extends a pretrained MDLM into a UDLM with an intrinsic ability to revise its own erroneous predictions during decoding. - Soft Parallel Decoding: Uses interpolation between mask and token embeddings to propagate confidence priors from previous steps. Tech stack: Python Why it matters: DMax offers a significant improvement in parallel decoding efficiency while maintaining high generation quality, making it ideal for dLLMs that require high performance and accuracy in parallel settings. Repository: czg1225/DMax GitHub stars: 132 GitHub description: [NeurIPS 2026] DMax: Aggressive Parallel Decoding for dLLMs