Projects

### Overview: HilbertRaum is a private AI workspace that runs locally on your computer and can operate entirely offline. Chat with a local AI, ask questions about your documents, and keep everything on your own computer. ### Key Features: - 🔒 **Private by design.** No cloud, no telemetry, no analytics. Nothing you type or import is uploaded. - 🧳 **Portable and encrypted.** Keep models and a password-encrypted workspace on an external drive and move between computers. - 🧠 **Local models.** A `llama.cpp` runtime with GGUF models from a curated open-weight catalog (Qwen3, Ministral, Gemma, Granite). The app benchmarks your machine and recommends one. - 📄 **Document Q&A with citations.** Import PDFs, Word files, or text, ask questions, and get answers grounded in your files. Retrieval combines vector and keyword search with a reranker and can be scoped to your library, a project, a section, specific documents, or the files attached to a chat. - 📚 **Knowledge packs — an offline Wikipedia (optional).** Register ZIM archives — Wikipedia in ~100 languages, Wiktionary, Wikivoyage — as per-chat sources searched and cited alongside your documents, with the article readable offline in the app. Needs the kiwix-tools binaries (GPL-3.0-or-later) on the drive — the app offers to install them with your consent, the first time you need them, or you can place them yourself

Overview: This project implements DreamBooth, a method for improving the quality of images, based on Stable Diffusion, with minimal code. Key features: - Implementation of DreamBooth using Stable Diffusion - Minimal code required for development and training Tech stack: Python Why it matters: This project provides a practical example of how to enhance image quality using a specific technique, demonstrating the integration of Stable Diffusion with DreamBooth for improved results.

### Overview: Arkor is a TypeScript framework designed for fine-tuning open-weight language models. It provides a type-safe configuration, local Studio, and managed GPUs, making it easy to ship custom models the same way TypeScript applications are developed. ### Key features: - **Fine-tuning from a TypeScript file:** `createTrainer({ model, dataset, lora, ... })` for LoRA training. - **Three curated templates:** `triage`, `translate`, and `redaction` with Gemma 4 base models. - **Typed lifecycle callbacks:** `onStarted`, `onLog`, `onCheckpoint`, `onCompleted`, and `onFailed` for event exposure. - **Checkpoint inference:** `infer({ messages })` for testing checkpoints during training. - **Local Studio:** `arkor dev` for a web UI with job list, loss chart, logs, and Playground. - **Anonymous workspaces and OAuth:** onboarding support for anonymous workspaces. ### Tech stack: TypeScript ### Why it matters: Arkor simplifies the process of fine-tuning language models by providing a type-safe configuration and a local development environment, making it easier for TypeScript developers to experiment and ship custom models.

### Overview: This repository contains a curated collection of papers and resources on the latest research in the field of Mixture of Experts (MoE) and related topics. The survey aims to provide a comprehensive overview of the evolution and advancements in MoE, including its applications in computer vision, large language models, specialized experts networks, and enhancing system performance and efficiency. ### Key features: - **Sparse Mixture of Experts (MoE-FFD)**: A paper that introduces a method for generalized and parameter-efficient face forgery detection. - **MoE in Visual Domain**: Papers exploring MoE in various visual tasks, such as face forgery detection and BEV detection. - **Large Language Models (LLMs)**: Papers on the application of MoE in large language models. - **Specialized Experts Network**: Papers focusing on specialized experts networks within MoE. - **Enhancing System Performance and Efficiency**: Papers discussing the use of MoE to improve system performance and efficiency. - **MoE in Recommendation**: Papers exploring the application of MoE in recommendation systems. - **Python Libraries**: A collection of Python libraries that can be used to implement MoE algorithms. ### Tech stack: Python ### Why it matters: This survey provides a valuable resource for researchers and practitioners interested in the latest advancements in MoE and related fields. It serves as a comprehensive guide to understanding the evolution of MoE and its applications, making it an essential tool for anyone looking to stay updated in this rapidly developing field.

Overview: This repository contains the code from the book "Generative AI in Action," published by Manning. It includes examples from the book, research papers related to AI technologies, a local web application, and detailed installation instructions. Key features: - Code from the book's examples - List of research papers - Local web application demonstrating various AI concepts - Detailed installation instructions Tech stack: Python Why it matters: The book's code works with the latest API (v1.0) from OpenAI, which is important for developers and data scientists who want to stay updated with the latest advancements in generative AI. Repository: bahree/GenAIBook GitHub stars: 146

Overview: This project implements a minimal version of a denoising diffusion probabilistic model (DDPM) for MNIST digits, achieving a model size of 4.55MB, making it ultra lightweight. Key features: - A fully functional denoising diffusion probabilistic model - Uses simple depthwise convolutions, shortcuts, and naive timestep embedding - Achieves 60 epochs of training in 12 minutes Tech stack: Python Why it matters: This project showcases the practicality and efficiency of implementing a complex model from scratch, demonstrating the importance of lightweight and efficient models in practical applications.

### Overview: LatentUnfold is a project implementing the paper "Flux Already Knows - Activating Subject-Driven Image Generation without Training". This project uses a vanilla Flux model to achieve subject-driven image generation, leveraging grid-based image completion and identity-preserving capabilities without additional data, training, or fine-tuning. ### Key features: - **Subject-driven image generation**: Activates strong identity-preserving capabilities using simple grid-based image completion. - **Free lunch approach**: Replicates subject images in a mosaic layout to activate identity-preserving capabilities. - **Cascade attention design**: Enhances identity preservation with a novel cascade attention design. - **Meta prompting technique**: Boosts fidelity and versatility with a meta prompting technique. - **Versatile edits**: Supports diverse edits such as logo insertion, virtual try-on, and subject replacement or insertion. ### Tech stack: Python ### Why it matters: Our method outperforms baselines across multiple key metrics in benchmarks and human preference studies, demonstrating that a pre-trained foundational text-to-image model can enable high-quality, resource-efficient subject-driven generation, opening new possibilities for lightweight customization in downstream applications.

### Overview: This repository is a curated list of public datasets used for object detection and recognition, covering various domains such as autonomous driving, adverse weather conditions, person detection, and more. It serves as a valuable resource for developers and researchers looking to enhance their projects with high-quality datasets. ### Key Features: - **Datasets Share Platform**: A platform for sharing datasets related to object detection and recognition. - **Datasets Tools**: A collection of tools and resources for data annotation, augmentation, and management. - **Datasets Tools (continued)**: - **Data Annotation**: Tools for labeling and annotating datasets. - **Data Augmentation**: Techniques to enhance dataset diversity. - **Data Management**: Strategies for organizing and managing large datasets. ### Tech Stack: The repository uses Markdown, GitHub, and various data sources to compile and share datasets. ### Why It Matters: This repository is essential for anyone working on projects that require object detection and recognition, providing a comprehensive list of publicly available datasets that can be used to improve accuracy, diversity, and effectiveness of their models.

### Overview: The 2024 Conference on Computer Vision and Pattern Recognition (CVPR) received 11,532 valid paper submissions, with 2,719 accepted papers, resulting in an acceptance rate of approximately 23.6%. This repository is a curated list of papers, workshops, datasets, and challenges from CVPR 2024, showcasing the latest research in computer vision and pattern recognition. ### Key Features: - **Papers**: A collection of papers from CVPR 2024 that cover various topics in computer vision, including image and video analysis, object recognition, and more. - **Workshops**: Information about upcoming workshops and tutorials related to CVPR 2024. - **Datasets**: A list of datasets used in the conference, including the HoHo dataset for 3D reconstruction and the PVUW dataset for video semantic segmentation. - **Challenges**: Details on challenges such as the Agriculture-Vision Prize Challenge, the Building3D Challenge, and the SyntaGen Competition, which aim to foster innovation and practical applications in computer vision. ### Tech Stack: Python ### Why It Matters: This repository provides valuable insights into the cutting-edge research and challenges in computer vision, making it an essential resource for developers and researchers interested in the field. The curated list ensures that you stay updated with the latest developments and can easily find relevant papers, datasets, and challenges to enhance your projects and contribute to the field.

### Overview: This repository compiles real-world implementations of Large Language Models (LLMs) and AI agents in the financial services sector, focusing on retail and commercial banking, wealth management, capital markets, payments, and insurance. It showcases enterprise platforms, fine-tuned models, and integrated solutions from leading financial institutions worldwide. ### Key Features: - **Enterprise Platform**: Large-scale LLM deployments across multiple lines of business or the entire organization. - **Fine-Tuned Model**: Models trained on specific datasets to achieve deep domain expertise. - **Integrated Solution**: Vendor-provided or off-the-shelf AI solutions deployed for specific workflows. ### Tech Stack: LLM and AI agent implementations in financial services. ### Why It Matters: LLMs and AI agents are transforming financial services, streamlining regulatory compliance, accelerating deal analysis, personalizing customer experiences, and automating complex workflows. This repository provides a comprehensive overview of real-world applications, showcasing the impact of these technologies on the industry.

Overview: The PALI-3 model is an implementation of a vision-language model designed for efficient and powerful processing. It leverages a contrastively pretrained SigLIP vision model to encode images into visual tokens, which are then processed through a 3B UL2 Transformer to generate text outputs. Key features: - Contrastive pretrained SigLIP vision model for image encoding - 3B UL2 Transformer for text generation - Efficient processing and strong performance Tech stack: Python Why it matters: The PALI-3 model is designed to be both efficient and powerful, making it suitable for a wide range of applications requiring fast and accurate language processing from images. Repository: kyegomez/PALI3 GitHub stars: 146 GitHub description: Implementation of PALI3 from the paper PALI-3 VISION LANGUAGE MODELS: SMALLER, FASTER, STRONGER README excerpt: ```python import torch from pali3.main import Pali3 model = Pali3() img = torch.randn(1, 3, 256, 256) prompt = torch.randint(0, 256, (1, 1024)) mask = torch.ones(1, 1024).bool() output_text = torch.randint(0, 256, (1, 1024)) result = model.process(img, prompt, output_text, mask) print(result) ```

Overview: Aegis Stack is a production-ready FastAPI platform that provides a complete set of components for building full-stack Python applications, including API, workers, scheduler, database, authentication, and AI integration. It also includes a built-in control plane for monitoring and managing the system. Key features: - Full-stack Python apps that know what they are - Assembles a complete FastAPI system with modular components - Built-in control plane for drawing, explaining, and catching changes - Supports private finance systems, meeting agents, MCP servers, local AI apps, and voice-first assistants - Provides a live map of the system, from the whole stack down to a single function - One command or one button in Overseer to provision a server in your Hetzner or DigitalOcean account - A live map of the system, with errors, deploys, jobs, resources, and cost - A live operator that reads the map, answers with diagrams, and reviews pull requests - Supports multiple languages for the CLI, including English, German, Spanish, French, Japanese, Korean, Russian, Simplified Chinese, and Traditional Chinese Tech stack: Python, FastAPI, HTMX, Pydantic, PostgreSQL, and others Why it matters: Aegis Stack provides a complete and production-ready platform for building full-stack Python applications, making it easier and faster to build, deploy, and manage applications with built-in monitoring and control features.

Overview: 这个项目是一个基于Taro + Python FastAPI + LangChain + LangGraph的AI闯关学习小程序，旨在通过AI编程的方式从零开始开发一个微信小程序，并覆盖从需求调研到部署上线的全流程，帮助用户快速掌握AI编程的完整工作流。 Key features: - AI出题和闯关答题功能 - 实时讲解和错误反馈 - AI生成的通关复盘报告 - 支持微信登录和微信小程序功能 - 知识库接入和RAG技术 - 上传私有文档功能 - AI生成的图片和对象存储 - Docker部署和微信云托管 Tech stack: Python, Taro, FastAPI, LangChain, LangGraph, DeepSeek, MySQL, JWT, Tencent Cloud COS, Alibaba Cloud, RAG, LangGraph, LangChain, Taro, FastAPI, LangChain, LangGraph, DeepSeek, MySQL, JWT, Tencent Cloud COS, Alibaba Cloud, RAG, LangGraph, LangChain, Taro, FastAPI, LangChain, LangGraph, DeepSeek, MySQL, JWT, Tencent Cloud COS, Alibaba Cloud, RAG, LangGraph, LangChain, Taro, FastAPI, LangChain, LangGraph, DeepSeek, MySQL, JWT, Tencent Cloud COS, Alibaba Cloud, RAG, LangGraph, LangChain, Taro, FastAPI, LangChain, LangGraph, DeepSeek, MySQL, JWT, Tencent Cloud COS, Alibaba Cloud, RAG, LangGraph, LangChain, Taro,

### Overview: These 211 free visual cards provide a comprehensive guide to understanding large language models (LLMs). Each card includes a diagram and explanation, making it easy to prepare for interviews, revisit learned concepts, or understand the systems being built. ### Key Features: - **Tokenization**: Understanding how sentences are broken down into tokens. - **KV Cache**: Reusing keys and values to improve performance in LLMs. - **Self-attention**: Mechanism for LLMs to focus on relevant information. - **Multi-head attention**: Enhancing the ability of LLMs to understand context. - **Agents**: Models that interact with tools and environments. - **RAG (Relevance-Augmented Generation)**: Retrieval-Augmented Generation for better results. - **Autoregressive generation**: Generating sequences based on previous inputs. - **Grouped-query attention**: Improving the accuracy of LLMs by grouping queries. ### Tech Stack: - Markdown ### Why It Matters: These cards serve as a valuable resource for developers and researchers looking to enhance their understanding of LLMs, from tokenization to deployment and beyond.

### Overview: This beginner-friendly series introduces Generative and Agentic AI through fun doodles and practical coding projects. Perfect for those curious about the basics of AI and how it connects to real-life applications. ### Key features: - **Generative AI and Agentic AI Explained**: Dive into the fundamental concepts of Generative AI and Agentic AI, learning how they differ from previous generations of AI. - **Hands-on Projects**: Build AI-powered apps using TypeScript, Azure AI Foundry, and other tools, including a Teams bot for Microsoft 365. - **GitHub Copilot and Azure OpenAI Services**: Utilize GitHub Copilot for Vibe-coding and Azure OpenAI services for advanced AI development. ### Tech stack: TypeScript, GitHub Copilot, Azure AI Foundry, Microsoft 365. ### Why it matters: Explore the basics of Generative and Agentic AI through fun doodles and practical coding projects, making the concepts accessible and relatable. Perfect for beginners looking to enhance their skills in AI development.

Overview: Generative UI represents a paradigm shift from "AI assists coding" to "AI generates interfaces directly." Instead of writing components, you describe what you need and the AI produces a working UI. Key features: - Dynamic UI generation - Personalization at scale - Rapid prototyping - Natural language interfaces Tech stack: AG-UI, A2UI, OpenUI Lang, MCP Apps, OpenAI's Apps SDK Why it matters: Enables personalization, rapid delivery, and natural interfaces, while reducing development time.

### Overview: Spring AI Playground is a cross-platform desktop app designed to safely execute and validate MCP tools. It provides a controlled local environment for building, testing, and deploying tools, ensuring they are secure and easy to inspect before reuse. ### Key Features: - **Safe Execution**: Every tool built earns a "Local Pass" before being added to the built-in MCP server, ensuring only safe tools are live. - **Observability Dashboards**: All tool interactions and calls are logged in built-in dashboards, including traces, tokens, cost, AI models, and more. - **Default Tools Catalog**: A bundled catalog of default tools for common use cases, searchable and filterable. - **External MCP Server Presets**: A preset catalog of external MCP servers grouped by category, with placeholders for environment variables. ### Tech Stack: Java ### Why It Matters: - Ensures that only safe tools are executed, reducing the risk of unintended consequences. - Provides a controlled environment for building and testing tools, making them easier to inspect and reuse. - Logs all interactions and calls, enabling detailed visibility into tool usage and performance. - Bundles default tools for common use cases and external MCP servers for more specialized environments. ### Repository: spring-ai-community/spring-ai-playground ### GitHub Stars: 145

### Overview: ReForm-Eval is a benchmark designed to evaluate the capabilities of large vision-language models (LVLMs) by reformulating existing task-oriented benchmarks into unified formats. This approach leverages existing annotations to provide substantial data for assessing LVLMs' freeform text outputs, facilitating a more comprehensive evaluation of their capabilities. ### Key Features: - **Unified Re-formulation**: Re-formulates existing benchmarks into a unified format that is compatible with LVLMs. - **Comprehensive Data**: Provides substantial data for evaluating various capabilities of LVLMs. - **Open Source**: The benchmark and evaluation framework will be open-sourced as a cornerstone for advancing the development of LVLMs. ### Tech Stack: Python ### Why It Matters: ReForm-Eval offers a novel approach to evaluating LVLMs by leveraging existing benchmarks and annotations, providing a more comprehensive and data-driven evaluation of their capabilities.

Overview: HiGPT is a heterogeneous graph language model that leverages in-context heterogeneous graph tokenization and instruction tuning for enhanced performance across various graph scenarios. Key features: - In-context heterogeneous graph tokenizer - Heterogeneous graph instruction-tuning framework Tech stack: Python Why it matters: HiGPT offers a unified approach to aligning large language models with heterogeneous graph structural knowledge, improving their ability to understand and utilize diverse graph instructions effectively.

### Overview: OpenOmni is an end-to-end fully open-sourced method that integrates image, speech, and text into an omni large language model. It achieves real-time emotional speech synthesis, enabling quick training with limited omni-modal data and VRAM resources. The framework supports two real-time modes: CTC mode and AR mode, allowing users to balance generation speed and quality based on their needs. OpenOmni's design makes it highly versatile and easy to apply to various downstream tasks, such as speech embodied navigation and multi-role-playing speech dialogue. ### Key Features: - **Real-Time Emotional Speech Synthesis**: Supports CTC mode and AR mode for flexible real-time speech generation. - **Image, Speech, and Text Integration**: Seamlessly integrates image, speech, and text into a unified omni model. - **Pioneering Method**: The first fully open-sourced approach for advancing omni-modal large language models. - **Versatile Framework**: Flexible framework design allows easy and quick application to various tasks. ### Tech Stack: Python ### Why It Matters: OpenOmni is a groundbreaking method that opens up new possibilities in the field of omni-modal large language models, making it easier and faster to apply to diverse applications. Its real-time emotional speech synthesis capabilities and integration of multiple modalities make it highly versatile and practical for various tasks.