Projects

### Overview: Space is a unified storage solution for the entire machine learning lifecycle, providing ground truth databases, OLAP databases, and distributed data processing pipelines. ### Key features: - **Ground Truth Database**: Manages multimodal data in open source file formats, supports data manipulation and version control. - **OLAP Database and Lakehouse**: Optimized for unstructured data, supports reference operations and SQL engines. - **Distributed Data Processing Pipelines**: Integrates with processing frameworks for efficient data transformation and materialized views. ### Tech stack: Python ### Why it matters: Space simplifies data management and storage across the entire machine learning lifecycle, making it easier to integrate with various ML frameworks and datasets.

### Overview: d3LLM is an ultra-fast diffusion language model that achieves 5x speedup over autoregressive models on high-performance GPUs, making it highly efficient for real-time applications. ### Key Features: - **Ultra-Fast Performance:** d3LLM offers up to 5x speedup over autoregressive models on H100 GPUs and 3.6x speedup on A100 GPUs, as well as 10x speedup over the vanilla Dream/LLaDA. - **Pseudo-Trajectory Distillation:** The model distills pseudo-trajectory information to achieve negligible accuracy degradation, making it highly accurate while being fast. - **Efficient Serving with SGLang Engine:** The models are now supported in the SGLang engine, enabling efficient serving with the latest GPU hardware. ### Tech Stack: The project uses Python, transformers, lm_eval, datasets, and flash_attn. ### Why It Matters: d3LLM is designed to be highly efficient and accurate, making it suitable for real-time applications and real-time generation tasks.

### Overview: Awesome AGV is a comprehensive set of standards and practices designed to elevate the capabilities of AI coding agents. It provides strict rules and specialized skills to ensure secure, defensible, and maintainable code. ### Key Features: - 📏 **25 Rules** — covering security, reliability, architecture, and maintainability. - 🛠️ **68 Skills** — specialized capabilities loaded on demand. - ⚡ **11 Command Skills** — end-to-end development processes. - 🤖 **21 Agent Personas** — specialized sub-agents for multi-agent orchestration. - 🏗️ **Three-Tier Loading System** — always-on mandates and on-demand skills. ### Tech Stack: Shell ### Why it Matters: While originally designed for Antigravity, Awesome AGV is built on standard markdown-based context protocols, making it easily portable to other AI coding tools. The principles of the Rugged Software Constitution, based on the Rugged Software Manifesto, improve the output of any LLM-based coding assistant.

Overview: NeuroLink is the pipe layer of an AI nervous system, connecting providers' neurons to application connectors via three inference types: generate, stream, and decide. It powers Tara, Yama, and Clairvoyance at Juspay. Key features: - **NeuroLink is the pipe layer of an AI nervous system.** - **It connects providers' neurons to application connectors via three inference types: generate, stream, and decide.** - **It powers Tara, Yama, and Clairvoyance at Juspay.** Tech stack: TypeScript Why it matters: NeuroLink provides a practical, TypeScript-first way to plug any application into the AI nervous system, offering three genuine inference types and a curated model registry, making it a versatile and efficient solution for integrating AI intelligence into applications.

Overview: This project profile outlines a comprehensive course on utilizing LangChain for enhancing the capabilities of language models in application development. Key features: - Models, Prompts, and Parsers - Memories for LLMs - Chains - Question Answering over Documents - Agents Tech stack: Jupyter Notebook Why it matters: This course provides essential skills to effectively leverage LangChain, enabling developers to enhance the use cases and capabilities of language models in various applications.

### Overview: This repository contains a collection of notes and resources related to large language models (LLMs). It serves as a comprehensive guide for developers and researchers interested in understanding and utilizing LLMs effectively. ### Key Features: - **Transformer Models**: A detailed introduction to the Transformer architecture, including its applications in translation and summarization. - **BERT and BART**: Information on the BERT and BART models, focusing on their improvements and applications. - **LLM Year Classification**: A table that categorizes known LLM models by their release year, providing a historical overview. - **LLM Implementation**: Resources and implementations for various LLM models, including TensorFlow, PyTorch, and others. ### Tech Stack: Jupyter Notebook The primary language used in this repository is Jupyter Notebook, which is an interactive computing environment for data science and machine learning. ### Why It Matters: This repository is essential for developers and researchers looking to understand and utilize LLMs effectively. It provides a comprehensive overview of various LLM models, their implementations, and their applications, making it a valuable resource for anyone interested in the field of natural language processing and artificial intelligence. ### GitHub Stars: 157 The repository has gained 157 stars from the developer community, indicating its popularity and usefulness in the field of LLMs.

Overview: The **Simple Chat Application** is a web-based platform designed for secure and context-aware interactions with generative AI models, utilizing Azure OpenAI for advanced AI capabilities. It features a modular design with optional features like Azure AI Content Safety, Image Generation, Video Indexer, and more, all secured with Azure Active Directory. Key features: - Secure and context-aware interactions with generative AI models. - Retrieval-Augmented Generation (RAG) for enhanced AI interactions. - Personal ("Your Workspace") and shared group ("Group Workspaces") document processing. - Integration with Azure AI Document Intelligence, Embeddings, Search, and Safety features. - Optional features like Azure AI Content Safety, Image Generation, Video Indexer, and more. - Azure Cosmos DB for conversation, metadata, and settings storage. - Azure App Service for deployment, supporting both Azure Commercial and Government environments. - Community Call for regular updates and discussions. Tech stack: Python, Azure OpenAI, Azure AI Document Intelligence, Azure AI Embeddings, Azure AI Search, Azure AI Content Safety, Azure Cosmos DB, Azure App Service, Azure Active Directory, Azure App Service, Azure Cosmos DB, Azure App Service, Azure Cosmos DB, Azure App Service, Azure Cosmos DB, Azure App Service, Azure Cosmos DB, Azure App Service, Azure Cosmos DB, Azure App Service, Azure Cosmos DB, Azure App Service, Azure Cosmos DB, Azure App Service, Azure Cosmos DB, Azure App Service, Azure Cosmos DB, Azure App Service, Azure Cosmos DB, Azure App Service, Azure Cosmos DB, Azure App Service, Azure Cosmos

### Overview: Portable-Diffusion is a premium, zero-configuration local AI image generator and offline Stable Diffusion GUI. It leverages hardware-accelerated GPU and NPU execution on Windows, Linux, and macOS, ensuring 100% offline functionality without requiring internet access, telemetry, or cloud logging. The application is designed for ease of use and portability, with all runtime components self-contained, eliminating the need for a global system environment change. ### Key Features: 1. **100% Offline & Private:** Runs inferences locally, requiring no internet, telemetry, cloud logging, or API keys. 2. **Zero-Install Portability:** Entire runtime is self-contained, ensuring no global system environment changes. 3. **Auto-Configured Acceleration:** Detects hardware specs to load CUDA (Nvidia), ROCm (AMD), Vulkan (Intel/AMD), Metal (macOS), or OpenVINO (Intel NPU) backends. 4. **Integrated Model Manager:** Supports pasting Hugging Face URLs to download weights directly, or drag-and-drop local `.safetensors` or `.gguf` files. 5. **Live Performance Monitor:** Tracks CPU, RAM, GPU, and VRAM utilization in real-time directly inside the web UI. 6. **Local Output Gallery:** Saves generated images side-by-side with prompt parameters and metadata JSON files. ### Tech Stack: - JavaScript ### Why It Matters: Portable-Diffusion offers a seamless, efficient, and private way to generate AI-generated images without the need for an internet

### Overview: Living Memory Coordinate-5 (LMC-5) is a five-axis memory architecture designed for long-running Large Language Models (LLMs). It offers a minimal SQLite + production PostgreSQL/pgvector reference implementation, focusing on multi-channel recall pipelines, LLM-proposed hippocampus, persona policies, and VPS 7×24 deployment shapes. ### Key Features: - **Minimal SQLite + PostgreSQL/pgvector**: A lightweight, off-the-shelf memory system. - **Multi-Channel Recall Pipeline**: Enhanced recall capabilities with independent channels. - **LLM-Proposed Hippocampus**: A more sophisticated memory structure inspired by the hippocampus. - **Persona Policies**: Personalized policies to manage user interactions and preferences. - **VPS 7×24 Deployment Shape**: Optimized for high availability and scalability. ### Tech Stack: - Python ### Why It Matters: LMC-5 is a compact, offline-first memory architecture tailored for LLM agents. It avoids the need for an infinite prompt, focusing on continuous memory recovery rather than infinite context. This design is particularly useful for agents that require long-term memory without being tied to a single model provider.

### Overview: This project is a comprehensive AI full-stack development initiative led by a Java backend engineer. The goal is to build a self-hostable RAG (Relevance-Augmented Knowledge) knowledge base. The project utilizes various technologies including Python, FastAPI, LangChain, PostgreSQL, pgvector, Next.js, and Docker. The primary focus is on creating a robust and scalable RAG pipeline that can handle document uploads, question-answering, and source citation retrieval. ### Key Features: - **Frontend**: Built with Next.js, it includes features for uploading documents, managing the knowledge base, and interacting with the source citations. - **Backend (Python)**: Utilizes FastAPI for the API, LangChain for the RAG pipeline, and PostgreSQL for storage and vector search. The project also leverages pgvector for efficient vector search. - **Documentation**: Provides a detailed roadmap and learning log, covering the entire project from week 1 to week 8. ### Tech Stack: - **Python**: FastAPI for the backend, LangChain for the RAG pipeline. - **FastAPI**: For building the API. - **LangChain**: For the RAG pipeline. - **PostgreSQL**: For storing and searching documents. - **pgvector**: For efficient vector search. - **Next.js**: For the frontend. - **Docker**: For deployment and CI/CD. ### Why It Matters: This project serves as a practical demonstration of how to build a self-hostable AI knowledge base from scratch, integrating various AI technologies. It showcases the ability to build a

Overview: PDFStract is a unified data preparation layer designed for RAG pipelines, offering flexibility in extracting, chunking, and embedding text from PDFs. It supports multiple libraries, methods, and providers, allowing developers to easily switch their data preparation strategy without code changes. Key features: - Extract: 10+ libraries (Marker, Docling, PyMuPDF4LLM, PaddleOCR, Unstructured, and more) - Chunk: 10+ methods (Token, Semantic, Sentence, Recursive, Code-aware, and more) - Embed: Multiple providers (OpenAI, Azure, Google, Ollama, Sentence Transformers) Tech stack: Python Why it matters: PDFStract simplifies the complex process of preparing data for RAG pipelines by providing a single API that can switch between various extraction, chunking, and embedding tools seamlessly. This flexibility allows developers to focus on their RAG outcomes rather than managing multiple library dependencies, making the process more efficient and easier to manage.

Overview: DiffPortrait3D is an official repository of a conditional diffusion model designed for zero-shot portrait view synthesis, capable of generating 3D-consistent photo-realistic novel views from a single in-the-wild portrait. Key features: - Zero-shot portrait view synthesis - Synthesizing plausible but consistent facial details from novel camera views - Retaining identity and facial expression - Generalizing well to unposed camera views, extreme facial expressions, and diverse artistic depictions Tech stack: Python Why it matters: This project demonstrates the potential of conditional diffusion models in generating realistic and consistent 3D views from single in-the-wild portraits, addressing the challenges of zero-shot and arbitrary face portraits.

### Overview: CC-Pan is an efficient pan-sharpening framework that leverages channel-wise compression and diffusion techniques to generate high-resolution multispectral imagery from low-resolution inputs. This project is based on the paper "CC-Pan: Channel-wise Compression based Diffusion for Efficient Pan-Sharpening," which was published on arXiv and has been cited by several researchers. ### Key Features: - **Channel-wise Compression**: Compresses multispectral channels into a compact latent space. - **Diffusion Model Fine-Tuning**: Fine-tunes a pre-trained diffusion model to generate high-resolution images. - **Dual-Branch Adapter**: Enhances the model's performance through a lightweight adapter. - **Training Pipeline**: Includes a pretraining stage with a 1-channel Band-VAE followed by a latent diffusion stage. - **Offline Inference**: Provides offline inference entry points for quick evaluation. ### Tech Stack: - Python 3.10.12 - PyTorch 2.10 ### Why It Matters: CC-Pan offers a novel approach to pan-sharpening that significantly reduces computational costs and improves the quality of the output images, making it particularly useful in applications where high-resolution imagery is required but the available resources are limited. ### Repository: The repository contains the training pipeline, offline inference entry points, checkpoint layout, and a Gradio demo to reproduce the paper workflow. It is hosted on GitHub and has been starred by 156 developers. ### At a Glance: - **Task**: Pan-sharpening from a

### Overview: This repository demonstrates how to create a simple agent using Guidance and local LLMs, providing a comprehensive guide on integrating Guidance with various Huggingface LLMs. The project showcases the power of local LLMs and Guidance in controlling and guiding the agent's responses. ### Key Features: - **Local LLMs Integration**: Utilizes Guidance to create a simple agent that can interact with local LLMs. - **Guidance Tool**: A tool for controlling LLMs, which can be used to build prompt templates. - **Gradio UI**: Added for a more user-friendly interface, allowing the agent to be run as a web application. ### Tech Stack: Python ### Why It Matters: This project serves as a practical guide for developers looking to integrate local LLMs with Guidance, making it easier to create agents that can interact with a variety of models. The addition of a Gradio UI further enhances the usability of the agent, making it accessible to a broader audience. ### Repository: QuangBK/localLLM_guidance ### Primary Language: Jupyter Notebook ### GitHub Stars: 156 ### GitHub Description: Local LLM ReAct Agent with Guidance ### README Excerpt: The [Guidance](https://github.com/microsoft/guidance) tool is used to control LLMs, providing a concept for building prompt templates. This repository demonstrates how to create a simple agent using Guidance, integrating it with various Huggingface LLMs. The agent itself is built using Guidance, and

### Overview: GenCompositor is a generative video compositing framework that leverages diffusion transformers to automate the process of video editing. This project aims to streamline the production of video productions by integrating live-action footage, making it accessible and cost-effective. ### Key Features: - **Generative Video Compositing**: Automatically injects identity and motion information of foreground video into the target video. - **Diffusion Transformers (DiT)**: A novel pipeline based on diffusion models to preserve the consistency of the target video before and after editing. - **Interactive Customization**: Users can easily adjust the size, motion trajectory, and other attributes of dynamic elements added in the final video. ### Tech Stack: - **Python**: The primary programming language used for implementation. ### Why It Matters: - **Efficiency**: Reduces the need for extensive labor and expert collaboration, making video production more accessible and cost-effective. - **Automation**: Automates the tedious process of video editing, allowing for faster turnaround times. - **Consistency**: Ensures the quality and consistency of the final video by maintaining the background before and after editing. ### Repository: - **Repository**: [TencentARC/GenCompositor](https://github.com/TencentARC/GenCompositor) - **GitHub Stars**: 156 - **GitHub Description**: "ICLR 2026] GenCompositor: Generative Video Compositing with Diffusion Transformer" ### Contact: - **Authors**: - Shuzhou Yang - Xiaoyu Li - Xiaodong

### Overview: ST-LLM is an official implementation of the paper "ST-LLM: Large Language Models Are Effective Temporal Learners", focusing on temporal understanding and effective video question answering. ### Key features: - Joint spatial-temporal modeling within large language models for effective video understanding. - Dynamic masking strategy for improved performance. - Zero-shot video question answering on ActivityNet, MSRVTT, and MSVD datasets. ### Tech stack: Python ### Why it matters: ST-LLM demonstrates the effectiveness of large language models in temporal learning, providing valuable insights into how these models can be applied to improve video understanding and answer questions.

### Overview: This project is an official implementation of a NeurIPS 2024 paper that explores the application of graph learning in improving planning capabilities of large language models (LLMs). The paper aims to demonstrate that biases in attention mechanisms and auto-regressive loss can hinder LLMs' performance in graph-based decision-making tasks. ### Key features: - **Graph Learning for Planning**: The project introduces an additional Graph Neural Network (GNN) specifically designed to retrieve sub-tasks from a graph, enhancing the ability of LLMs to plan tasks effectively. - **Training-Free Methods**: The implementation includes both training-free and training-based variants of the GNN, which can be used without the need for additional training data. - **Benchmarking**: The project showcases the performance improvement of the proposed method across various LLMs and planning benchmarks, demonstrating significant computational efficiency. ### Tech stack: Python ### Why it matters: This project is significant because it provides a practical solution to improve the planning capabilities of LLMs by leveraging graph learning, which can lead to more efficient and effective task management in AI-driven applications.

Overview: This project allows developers to save images in multiple formats such as AVIF, WebP, JPEG, and JPEG2k, with the option to customize folder and subfolder names and filenames. It also includes functionality to save job data and metadata. Key features: - Save images in different formats - Customize folder and subfolder names and filenames - Save job data and metadata Tech stack: Python Why it matters: This project is useful for developers who need to save and organize images in a structured manner, especially when working with ComfyUI or similar tools that support multiple image formats. Repository: audioscavenger/save-image-extended-comfyui Primary language: Python GitHub stars: 156

Overview: This project is a collection of two agents, the Granite Retrieval Agent and the Image Research Agent, designed for research and image analysis using IBM's Granite platform. Key features: - Granite Retrieval Agent: A general agentic RAG system for document and web retrieval using Autogen/AG2. - Image Research Agent: An image-based multi-agent research system using CrewAI with Granite Vision, optimized for low latency and small footprint. Tech stack: Python Why it matters: Granite 4 introduces a hybrid Mamba-2/Transformer architecture that targets lower memory use and faster inference, making it a strong fit for agentic RAG and function-calling workflows. It uses >70% lower memory and ~2× faster inference compared to comparable models, which helps these agents run locally or on modest GPUs with lower cost and latency. Models are Apache-2.0 licensed, ISO 42001 certified, and cryptographically signed for governance and security. Granite 4 is optimized for low-latency, small-footprint deployments—ideal for the Image Researcher’s quick tool calls and orchestration steps. The family emphasizes instruction following, tool calling, RAG, JSON output, multilingual dialog, and code (incl. FIM), aligning with both agents’ needs.

Overview: The Kubeflow SDK is a Python library designed to simplify the deployment of AI workloads on Kubernetes, providing a unified interface for interacting with various Kubeflow projects. Key features: - Unified Python API for interacting with multiple Kubeflow projects - Abstracts away Kubernetes complexity, making it easier to work with AI applications - Seamlessly scales any AI workload, from local to large-scale production clusters - Reduces friction between development and production environments, enabling rapid iteration - Supports local development without the need for a Kubernetes cluster Tech stack: Python Why it matters: The Kubeflow SDK simplifies the deployment of AI workloads on Kubernetes, making it easier to build, develop, and scale AI applications without the need for deep knowledge of Kubernetes.