Projects

Overview: Argus 是一个基于 RAG 架构的开源知识库平台，后端采用 Java 21 + Spring Boot + MyBatis-Plus + PostgreSQL/pgvector，前端采用 Vue 3 + TypeScript + Element Plus，AI 层基于 Spring AI Alibaba（通义千问）+ ReactAgent 图引擎，以 MinIO + Elasticsearch 为存储与检索引擎。 Key features: - 自动文档解析 → 切片 → 向量化 → 索引，打通从文件到知识的全链路 - 混合检索架构：向量语义检索 + 关键词全文检索，RRF 融合排序 - 证据评估（四级充分度） → LLM 生成 → 引用溯源 - AI Agent 对话引擎：双模式切换、工具编排、SSE 流式输出、短期记忆 - 三级角色权限、JWT 双令牌、BCrypt 密码加密、群组数据隔离、AOP 操作日志 Tech stack: Java 21 + Spring Boot + MyBatis-Plus + PostgreSQL/pgvector + Vue 3 + TypeScript + Element Plus + Spring AI Alibaba + ReactAgent + MinIO + Elasticsearch Why it matters: Argus 是一个融合 RAG 与 AI Agent 技术的企业级智能知识平台，能够解决 LLM 在垂直领域应用中的三大核心痛点：幻觉编造、知识割裂、无记忆对话。它通过自研的查询规划、RRF

Overview: Kõnele is an Android app designed to provide speech-to-text user interfaces for other apps, utilizing speech recognition technology. Key features: - A voice search panel for Android devices - A speech keyboard implementing the Input Method Editor API - Two open source speech recognition servers: Kaldi-GStreamer Server and Ruby-Pocketsphinx Server Tech stack: Java Why it matters: Kõnele aims to support Estonian speech recognition on the Android platform and offers grammar-based speech recognition for voice command applications, enhancing user interaction with other apps. Repository: Kaljurand/K6nele GitHub stars: 290 GitHub description: An Android app that offers speech-to-text user interfaces to other apps.

Overview: The Spezi LLM Swift Package offers essential tools for integrating Large Language Models (LLMs) into your Spezi-based applications. This package includes modules for local LLM execution, remote OpenAI-based LLMs, and LLMs running on Fog node resources within the local network. Key features: - Local LLM execution via the `SpeziLLMLocal` target - Remote LLMs using the `SpeziLLMOpenAI` target - Support for various LLMs including Hugging Face models like Llama2, Phi, Gemma, and DeepSeek-R1 Tech stack: Swift Why it matters: This package simplifies the integration of advanced AI capabilities into your Spezi-based applications, making it easier to leverage cutting-edge language models without the need for extensive development effort.

### Overview: BOND is a framework designed for open-domain named entity recognition using distant supervision. It leverages pre-trained BERT models to enhance the accuracy of open-domain NER tasks. ### Key features: - **Distant Supervision**: Uses weakly labeled data to improve model performance. - **BERT-Assisted**: Incorporates BERT for enhanced feature extraction. - **Open-Domain**: Suitable for various domains and languages. - **Datasets**: Provides five open-domain distantly/weakly labeled NER datasets. ### Tech stack: Python ### Why it matters: BOND is a significant advancement in open-domain NER, offering a robust solution for tasks that require weak supervision. Its integration of BERT and distant supervision significantly boosts performance compared to previous methods, making it a valuable tool for researchers and practitioners in the field.

### Overview: This project presents an official implementation of "Towards Realistic Scene Generation with LiDAR Diffusion Models," showcasing state-of-the-art techniques for generating realistic scenes using LiDAR diffusion models. ### Key features: - Pretrained autoencoders and LiDMs for various tasks. - Detailed study of autoencoder design and pretrained models. - Available conda environment for easy setup. - Comprehensive evaluation toolbox for standardizing metrics. - Model zoo with pretrained LiDM and autoencoders for different tasks. ### Tech stack: Python ### Why it matters: This project offers a practical and efficient solution for generating realistic LiDAR scenes, which is crucial for applications in autonomous vehicles, augmented reality, and other fields that require high-fidelity 3D models.

Overview: JoliGEN is an integrated framework for training custom generative AI image-to-image models, combining GAN, Diffusion, and Consistency models for real-world applications such as controlled image generation, augmented reality, dataset smart augmentation, and synthetic to real transforms. Key features: - Implements both GAN, Diffusion, and Consistency models for unpaired and paired image to image translation tasks, including domain and style adaptation with semantic conservation. - Supports real-world applications like Controled Image Generation, Augmented Reality, Dataset Smart Augmentation, and object insertion, as well as Synthetic to Real transforms. - Provides a fast and stable training experience with astonishing results. A server with REST API is available for simplified deployment and usage. - Offers a large scope of options and parameters, with detailed documentation on models, dataset formats, and data augmentation. - Supports various use cases, such as image inpainting, image translation while preserving the class, object insertion, and simulation to reality translation while preserving elements and metrics. Tech stack: Python. Why it matters: JoliGEN is a powerful tool for developers and researchers looking to create custom generative AI models for real-world applications, offering a comprehensive solution with a focus on practical results and ease of use. Repository: jolibrain/joliGEN Primary language: Python GitHub stars: 290 GitHub description: Generative AI Image and Video Toolset with GANs and Diffusion for Real-World Applications README excerpt: <div align="center"> <img src="imgs/joligen.svg" width="5

### Overview: Open Deep Wide Research is an agentic Retrieval-Augmented Generation (RAG) agent designed for flexibility and customization. It offers granular control over sources, depth, and breadth, making it suitable for various scenarios and reducing the complexity of enterprise-level RAG development. ### Key Features: - **Customizable Sources**: Connect custom data sources from internal knowledge bases to specialized APIs. - **Controlled Depth**: Adjust response time and reasoning depth. - **Breadth Control**: Manage information breadth across selected sources. ### Tech Stack: TypeScript ### Why It Matters: In 2025, traditional RAG systems are evolving towards more dynamic, agentic RAG approaches. The market urgently needs a single, open-source RAG agent that is MCP-compatible and allows developers to tailor performance, scope, and cost effectively. Open Deep Wide Research addresses these needs by providing granular control over core dimensions of agentic research, including sources, depth, and breadth, and predicting the cost of each response.

### Overview: VeriInk is a tool that turns model failures into reproducible handwriting data, specifically designed for OCR, multimodal grading, post-training, and regression evaluation. ### Key features: - **Deterministic Rendering**: The renderer records the seed, visible answer, target transcription, ink bounds, and file hash for each sample. - **Synthetic Hard Cases**: Covers crossed-out revisions, decimal arithmetic, larger-number operations, and mixed-operation comparisons. All six images are from the simulator and do not include real student images. - **Public Audit Set**: The demo generates 16 pages and 80 answer regions, including arithmetic, decimals, fill-in-the-blank questions, comparisons, unit conversion, intentional mistakes, and crossed-out revisions. The `manifest.json` links every image to its seed, labels, bounds, and SHA256. ### Tech stack: Python ### Why it matters: VeriInk turns model failures into reproducible handwriting data, making it easier to audit and evaluate model performance in OCR, multimodal grading, post-training, and regression evaluation. ### Repository: znxxzsy/VeriInk - **GitHub Stars**: 290 - **Primary Language**: Python - **GitHub Description**: Verifiable handwriting synthesis for failure-driven multimodal training and regression. - **GitHub URL**: [https://github.com/znxzsy/VeriInk](https://github.com/znxzsy/VeriInk) - **Public Release Boundary**: The repository contains synthetic demo data only, excluding real student

Overview: This project focuses on beamforming techniques for speech enhancement, including simple delaysum, Minimum Variance Distortionless Response (MVDR), and Conjugate Gradient Minimum Variance (CGMM-MVDR). Key features: - Simple delaysum - MVDR - CGMM-MVDR Tech stack: Python Why it matters: This project provides a simple yet effective solution for speech enhancement, making it useful for developers looking to improve audio quality in their applications.

### Overview: This project is an experimental sandbox designed to test out ideas related to running local Large Language Models (LLMs) with Ollama and Pydantic AI to perform agentic Retrieval-Augmented Generation (RAG) for answering questions based on your documents. The agent can decide when and how to search documents, unlike fixed RAG pipelines. The project uses Ollama to create embeddings with nomic-embed-text stored in LanceDB for vector search. ### Key Features: - Agentic Retrieval-Augmented Generation (RAG) - Local LLMs with Ollama - Vector database for embedding storage and retrieval - Type-safe agent framework with tool calling (Pydantic AI) - Reliable tool calling for agentic RAG - Benchmarking of models for RAG performance - Supported document formats: PDF, Word, PowerPoint, Excel, Markdown, HTML, CSV, JSON ### Tech Stack: - Python ### Why It Matters: The project is designed to explore and test the capabilities of local LLMs combined with Ollama and Pydantic AI, focusing on agentic RAG for efficient and reliable document-based question answering. The use of Ollama and Pydantic AI ensures a robust and efficient RAG system, while the vector database and supported document formats make it versatile for various use cases.

Overview: Miko is a real-time, interactive AI companion designed to live on your Windows desktop. It pairs local Large Language Models (via Ollama) with two first-class voice synthesis engines: Fish Audio S2.1 Pro and GPT-SoVITS v2pro. Key features: - Dual voice synthesis options: Fish Audio S2.1 Pro and GPT-SoVITS v2pro - Persistent human-like conversation memory - Smart-home IoT control - Real-time lip-sync and physical turntable hardware integration - Custom VRM avatar support Tech stack: C# Why it matters: Miko offers a unique blend of AI and voice technology, making it a powerful tool for enhancing user experience and interaction on the Windows desktop.

Overview: `pdfdeal` is a Python wrapper for the Doc2X API, designed to enhance PDF recall in RAG systems. It includes native text processing capabilities to improve PDF performance. Key features: - Integrates with Doc2X for document conversion and OCR. - Provides native text processing for better PDF recall in RAG applications. - Supports conversion of images and PDFs to Markdown format. - Offers tools for splitting and adding separators to Markdown documents. Tech stack: Python Why it matters: Simplifies PDF handling and enhances recall in RAG systems, improving overall performance and accessibility. Repository: NoEdgeAI/pdfdeal Primary language: Python GitHub stars: 288

### Overview: This repository contains a collection of prompts designed to be used with Multimodal Language Models, specifically focusing on GPT-4V and DALL-E3. These prompts aim to fully utilize the multi-modal capabilities of the models. ### Key Features: - **GPT-4V Prompts**: A set of prompts tailored for GPT-4V, utilizing its advanced capabilities to generate more coherent and contextually relevant responses. - **DALL-E3 Prompts**: Prompts for DALL-E3, leveraging the model's ability to create realistic and detailed images, enhancing creativity and realism in visual outputs. ### Tech Stack: This repository is written in Markdown and uses the GitHub Actions workflow to manage its CI/CD pipeline. ### Why It Matters: These prompts provide developers with a library of examples and ideas to enhance the functionality and creativity of their Multimodal Language Models, making them more versatile and useful in various applications.

### Overview: PMB is a local-first persistent memory system for AI coding agents, providing decisions, lessons, and facts in one SQLite file on your disk. It eliminates the need for cloud storage, API keys, and LLM calls, ensuring offline and multilingual functionality. ### Key Features: - **Local Workspace:** Decisions, lessons, and facts live in a local SQLite file on your disk. - **Offline Support:** Works without internet connection, ensuring you can work without an internet connection. - **Multilingual:** Supports multiple languages, making it accessible to a global audience. - **Automatic Search:** Memory is automatically searchable, allowing quick access to past decisions and lessons. - **No Cloud Storage:** No need for cloud storage, eliminating the risk of data loss. - **Model Upgrade:** Memory persists even if the model is upgraded or switched, ensuring your agent's knowledge remains current. ### Tech Stack: Python ### Why It Matters: PMB provides a local-first memory system that eliminates the need for cloud storage, API keys, and LLM calls, ensuring offline and multilingual functionality. This makes it a valuable tool for developers and AI agents who need to work without internet connectivity or rely on local, persistent memory.

### Overview: qKnow 是一个面向企业知识智能化与行业 AI 应用的开源 AI 智能体构建平台，提供知识图谱、知识库 RAG、Bot 构建和开箱即用的 AI 应用等核心能力，支持文档、结构化数据、业务知识和专家经验的统一接入与智能化沉淀，帮助企业快速构建知识中枢、智能问答平台、智能体开发平台和行业深度 AI 解决方案。 ### Key features: - 🧠 知识图谱 + RAG 双引擎 - 🤖 面向企业的 Agent 构建 - 🧩 可视化编排 - 📚 企业知识资产沉淀 - 🔌 开放模型接入 - 🏗️ 插件化应用扩展 - ☕ Java 技术栈友好 ### Tech stack: Java ### Why it matters: qKnow 为企业知识智能化与行业 AI 应用提供了标准化模块，支持文档、结构化数据、业务知识和专家经验的统一接入与智能化沉淀，帮助企业快速构建知识中枢、智能问答平台、智能体开发平台和行业深度 AI 解决方案。

### Overview: This repository is a curated collection of Deep Learning methods and Foundation Models (LLM, LM, FM) specifically designed for trajectory data management and mining. It serves as a valuable resource for researchers and practitioners in the field, providing comprehensive papers, codes, datasets, and toolkits. ### Key Features: - **Deep Learning Models for Trajectory Computing**: A wide range of deep learning models are reviewed and discussed, focusing on their applications in trajectory data management and mining. - **Taxonomy Framework**: A taxonomy framework is provided to categorize trajectory data management and mining methods. - **Public Datasets**: A list of public datasets used in trajectory computing research is included. - **Toolkits**: Various toolkits and libraries are mentioned for trajectory data processing and analysis. ### Tech Stack: - Python - TensorFlow - PyTorch - Scikit-learn - Matplotlib ### Why It Matters: This repository is essential for anyone looking to stay updated with the latest advancements in deep learning techniques for trajectory data management and mining. It provides a comprehensive overview of the field, making it a valuable resource for researchers and practitioners alike. ### Repository: - **Repository**: yoshall/Awesome-Trajectory-Computing - **Stars**: 288 - **Description**: A professional list of Deep Learning and Large (Language) Models (LLM, LLM, FM) for Trajectory Data Management and Mining. - **GitHub**: <https://github.com/yoshall/Awesome-Trajectory-Computing> - **Icon**:

### Overview: Story2Board is a training-free framework for generating expressive storyboards from natural language descriptions. It addresses the limitations of existing methods by focusing on visual storytelling aspects such as spatial composition and background evolution, achieving state-of-the-art results without requiring extensive fine-tuning. ### Key Features: - **Latent Panel Anchoring**: Preserves a shared character reference across panels to ensure identity consistency. - **Reciprocal Attention Value Mixing (RAVM)**: Softly blends visual features between token pairs with strong reciprocal attention to enhance coherence. ### Tech Stack: - Python 3.12 - CUDA 12.x ### Why It Matters: Story2Board produces more dynamic, coherent, and narratively engaging storyboards compared to existing baselines, making it a valuable tool for visual storytelling.

Overview: DeepEar is an open-source framework designed for deep research and financial signal tracking, transforming public opinion into actionable investment logic chains. Key features: - Multi-Agent Collaboration: Specialized agents for trend spotting, financial analysis, and report writing. - Agent Skill Ready: Plug-and-play skill for AI agents (Antigravity, OpenCode), turning any chat assistant into a financial researcher. - Interactive Fintech Dashboard: Modern React-based UI for real-time monitoring of agent workflows, logs, and structured results. - Logic Evolution Tracking: Active tracking of investment theses as new market news and price data arrive. - Signal-Based Comparison: Side-by-side analysis of signal changes, drift in sentiment, and logic updates between different runs. - 15+ Data Sources: Comprehensive coverage including Weibo, Cailian Press, Wall Street News, and more via `NewsToolkit`. - News-Aware Time-Series Model: Custom Kronos integration with a news-projection layer for predicting price shocks. - Visualized Reports & Logic Graphs: High-fidelity HTML reports featuring interactive Draw.io diagrams for logic transmission. - Hybrid RAG Engine: Combines BM25 (keyword) and Vector Search (semantic) for precise information retrieval. Tech stack: Python Why it matters: DeepEar simplifies complex financial analysis by automating the process of converting qualitative news into quantitative sentiment scores and signal indicators, making it easier for researchers and investors to derive actionable insights from unstructured data. Repository: GitHub repository: HKUSTDial/DeepEar Primary language: Python GitHub stars: 2

### Overview: InterDiff is an innovative framework that addresses the challenge of generating realistic 3D human-object interactions (HOIs) using physics-informed diffusion. The method aims to model dynamic objects with various shapes, capture whole-body motion, and ensure physically valid interactions, making it significantly more challenging than existing research. ### Key Features: - **Physics-Informed Diffusion**: InterDiff uses a diffusion model to encode the distribution of future human-object interactions. - **Interaction Correction**: A physics-informed predictor corrects denoised HOIs in a diffusion step to ensure physically valid interactions. - **Prior Knowledge Injection**: The method injects prior knowledge that interactions under reference with respect to contact points follow a simple pattern and are easily predictable. ### Tech Stack: - Python 3.7 ### Why It Matters: InterDiff is a significant advancement in the field of 3D human-object interaction synthesis, offering a more comprehensive and physically accurate approach to generating realistic HOIs. This method can be applied to various applications, including virtual reality, robotics, and interactive simulations, where precise and realistic interactions are crucial.

### Overview: This repository is a curated collection of battle-tested tools, frameworks, and best practices for building, scaling, and monitoring production-grade Retrieval-Augmented Generation (RAG) systems. It provides essential resources for developers looking to integrate RAG into their applications, ensuring high performance and reliability. ### Key Features: - **Production-Grade RAG Systems**: Tools and frameworks optimized for real-world deployment. - **Vector Databases**: High-performance storage solutions for RAG. - **Evaluation Tools**: Metrics and benchmarks for assessing RAG models. - **Observability Solutions**: Techniques for monitoring and debugging RAG systems. - **Case Studies**: Real-world applications of RAG, including semantic search, Q&A systems, and AI-powered assistants. ### Tech Stack: - Python ### Why It Matters: RAG systems are crucial for integrating external knowledge into large language models, enhancing their capabilities and making them more effective in various applications. This repository provides the necessary tools and practices to build, scale, and deploy RAG systems reliably and efficiently. ### Repository: - **Repository**: Yigtwxx/awesome-rag-production - **GitHub Stars**: 284 - **GitHub Description**: A curated list of battle-tested tools, frameworks, and best practices for building scalable, production-grade Retrieval-Augmented Generation (RAG) systems. ### Featured in: - **Awesome Generative AI**: Featured in the Awesome Generative AI list. - **Awesome Machine Learning**: Featured in the Awesome Machine Learning list. - **Awesome Production ML**: Featured in the Awesome Production ML list.