Projects

Overview: Personal-Graph is a Python library designed to help AI applications, particularly Large Language Models (LLMs), build and manage knowledge graphs. It aims to enhance working and long-term memory capabilities. Key features: - Can Be Fast: Built on libsql, a high-performance SQLite engine in Rust. - One DB per User: Integrates with Turso DB for private, user-specific knowledge graphs. - Natural Language Interfaces: Powered by Sqlite-vss and Instructor for natural language queries. - ML-Ready: Export data for machine learning libraries like Networkx and PyG. - Supports local execution with Ollama LLMs and Embedding Models. - Supports local db for storing graph and embeddings. Tech stack: Python, SQLite, Turso DB, libsql, Sqlite-vss, Instructor, Networkx, PyG, Ollama LLMs, Embedding Models, DuckDB, TransE, TransE, Query2Box. Why it matters: Personal-Graph simplifies the process of building and managing knowledge graphs, making it easier for AI systems to access and utilize information efficiently. This can significantly improve the performance and accuracy of AI applications by providing a more comprehensive understanding of the user's context and experiences.

### Overview: GARF: Learning Generalizable 3D Reassembly for Real-World Fractures is a research project that presents a method for 3D reassembly of fractures from real-world datasets. The project has been accepted at ICCV 2025 and achieved strong performance across extensive benchmarks. ### Key features: - **Generalizable Flow Matching-based 3D Reassembly**: The method is designed to be highly adaptable and can be applied to real-world fracture reassembly tasks. - **Training on 1.9 Million Fractures**: The project leverages a large dataset of fractures to train the model, ensuring robust generalization. - **Efficient Performance**: The code is concise and performs well, making it suitable for real-world applications. ### Tech stack: - Python ### Why it matters: - The project demonstrates the potential of generalizable 3D reassembly methods, which can be applied to various real-world applications, including medical imaging and robotics. - The use of a large dataset for training ensures that the model can handle diverse and real-world scenarios effectively. ### Repository: - **Repository**: ai4ce/GARF - **GitHub Stars**: 105 - **Primary Language**: Python - **GitHub Description**: [ICCV2025] GARF: Learning Generalizable 3D Reassembly for Real-World Fractures ### Project Structure and Config System: - The project includes a structured project directory and a configuration system to streamline the development process. ### Evaluation Performance: - The project achieves strong performance across

Overview: Local Code is a private AI agent for Mac, iPad, and iPhone, powered by a combination of a quantized model and a code harness. It runs on Apple Silicon, with the ability to switch between local and cloud models. The app includes Python, Node.js, and a shell, making it accessible and versatile. Key features: - Swift-based app with a modular design - Runs inference engine, coding harness, and agent runtime on device - Supports local and cloud models - Durable tasks and threads, resumable generation, and long-term context cache - Quantized checkpoints for DeepSeek 4.1 Flash 1017 and DeepSeek 4 Flash 0731 Tech stack: Swift, Python, Node.js, shell, and libraries from the repository Why it matters: Local Code provides a powerful, user-friendly AI solution for Mac, iPad, and iPhone, combining cutting-edge AI technology with a familiar development environment.

Overview: llama.go is a pure Go implementation of the LLaMA model designed for inference and educational purposes. It features a clean, structured codebase matching HuggingFace reference layouts, supports LLaMA 1, 2, and 3 architectures, and includes features like HF-aligned architecture, concurrent MHA, int8 quantization, and zero dependencies. Key features: - HF-aligned Architecture - Concurrent MHA - Int8 Quantization - Zero Dependencies Tech stack: Go Why it matters: This repository provides a straightforward, educational way to understand transformer internals by running LLaMA inference using only Go's standard library, making it ideal for learning and exploring transformer models without heavy dependencies. GitHub stars: 105 Repository: gitctrlx/llama.go

Overview: Crumbs is a tool that converts asterisk-indented text files into mind maps, making it easier to organize and visualize notes hierarchically. Key features: - Converts asterisk-indented text files into mind maps. - Supports different levels of indentation (1, 2, 3, etc.) for hierarchical organization. - Generates DOT scripts for Graphviz, allowing for easy visualization. Tech stack: Go Why it matters: Crumbs simplifies the process of organizing and visualizing notes by providing a straightforward way to convert text into mind maps, which can be easily visualized and shared.

Overview: llmaker is an open-source platform that allows developers to self-host modern language models, vector databases, and other components of the LLM stack from a single command. This eliminates the complexity of assembling and managing multiple services, making it easier to build and deploy applications locally without relying on third-party APIs or data transfers. Key features: - One command to provision and run the entire modern LLM stack. - OpenAI-compatible models. - Built-in vector database (Qdrant) and caching layer. - Observability and a built-in retrieval layer. - Easy to use and maintain. Tech stack: Go Why it matters: llmaker simplifies the process of setting up and managing modern LLM stacks, making it easier for developers to build and deploy applications locally without the need for third-party services or data transfers. This is particularly useful for developers who want to control their infrastructure and avoid the complexity of managing multiple services and dependencies.

Overview: Citrus is an open-source distributed vector database designed for efficient storage and retrieval of large amounts of vector data. Key features: - Distributed architecture for high availability and scalability. - Supports vector embeddings for fast similarity search. - Easy installation and usage via Python client. - Provides real-time updates and automatic reindexing. Tech stack: Python Why it matters: Citrus is ideal for applications requiring high performance and real-time updates, such as recommendation systems, image search, and virtual assistants. Repository: 0xDebabrata/citrus Primary language: Python GitHub stars: 104 GitHub description: (distributed) vector database

Overview: Arduinobot is an open-source 3D printed robot arm powered by ROS 2, designed for educational purposes and practical learning. It features a simple design and low cost, making it an excellent tool for robotics and ROS 2 learning. Key features: - `arduinobot_bringup`: Launch files for starting all functionalities of the real or simulated robot. - `arduinobot_controller`: ROS 2 Control configuration and hardware interface with the real robot. - `arduinobot_cpp_examples`: ROS 2 templates for developing with C++. - `arduinobot_description`: URDF description and Gazebo simulation. - `arduinobot_firmware`: Arduino code for actuating the motors of the real robot. - `arduinobot_moveit`: MoveIt 2 configuration and launch files. - `arduinobot_msgs`: Definition of new ROS 2 message interfaces. - `arduinobot_py_examples`: ROS 2 templates for developing with Python. - `arduinobot_remote`: Remote interface for moving the robot using Amazon Alexa voice assistant. - `arduinobot_utils`: Utilities and tools. Tech stack: C++ and ROS 2. Why it matters: Arduinobot is featured in the "Robotics and ROS 2 - Learn by Doing! Manipulators" course on Udemy, making it an excellent resource for learning robotics and ROS 2 through practical application.

### Overview: LodeDB is the world's fastest and most compact embedded vector database, designed for local RAG. It offers exact search by default, in-process, and on-disk, making it the ideal solution for LangChain, LlamaIndex, and mem0. With a minimal footprint and GPU-accelerated batched search, LodeDB is 13.6× smaller on disk, 600× faster per single query, and 2,880× faster per batched retrieval compared to the default stores. It also provides 26,000× faster add of one memory compared to other stores like SQLite-Vec, Qdrant, and Lancedb. ### Key Features: - **GPU-Resident Batch Search**: A float32 copy of the index lives on the GPU, scored with cuBLAS GEMM plus an on-device top-k. - **O(changed) Persistence**: Commits only the rows that changed, 173× to 1,308× faster than a full rewrite. - **Compact Storage**: The MIT TurboVec core packs vectors into 2/4-bit codes, and retained document text is stored zstd-compressed. - **In-process, on-disk**: No daemon, no account, no API key. - **Safe Concurrency**: One writer and many lock-free readers per path; every commit is crash-atomic and rolls back to the last committed state on failure. - **Private by Default**: Text, IDs, and vectors stay local

### Overview: The Higgsfield MCP Unified Server is a local-first Model Context Protocol (MCP) server designed for the Higgsfield AI API. It includes 85 models such as Seedance 2.5, Kling 3.0/O3/Omni, Wan 3, Genjutsu, and more. The server supports webhooks, cost estimates, and structured output, making it easy to integrate with various clients. ### Key Features: - **85 Models**: 85 models from the official Higgsfield API, including Seedance 2.5, Kling 3.0/O3/Omni, Wan 3, Genjutsu, and more. - **Webhooks**: Integrates with webhooks for seamless integration. - **Cost Estimates**: Prices requests without generating models. - **Structured Output**: Returns schema-based results instead of opaque blobs. - **Reliability**: Idempotent key handling and circuit breaker for retries. - **Webhooks and Outputs**: Supports images, video, audio, and additional video/3D files. ### Tech Stack: Python ### Why It Matters: The local-first approach ensures that every model's parameters are validated before any charges are made, providing a more secure and reliable experience. Additionally, it offers a cost-free discovery process and structured output, making it easier to integrate with various clients.

Overview: Deepmark AI is a benchmarking tool designed to assess large language models (LLM) on task-specific metrics and on your own data, ensuring predictable and reliable performance for GenAI-powered solutions. Key features: - Task-specific metrics assessment - Integration with leading Generative AI APIs - Pre-built for GPT-4, Anthropic, GPT-3.5 Turbo, Cohere, AI21, and others - Seamless assessment of LLMs on your own data Tech stack: PHP Why it matters: Deepmark AI provides a solution for organizations needing to assess LLM models on their own data, ensuring verifiable results and balanced accuracy, precision, recall, and reliability, addressing the challenges of GenAI and enabling seamless adoption in various domains.

Overview: This comprehensive repository provides a detailed guide on building, deploying, and managing intelligent AI agents and RAG pipelines using various advanced frameworks and tools. It covers LangChain, AutoGen, MCP, and n8n for automating AI workflows, and includes hands-on projects for fine-tuning models and deploying high-performance inference endpoints. Key features: - LangChain-Basics - AutoGen-Introduction - MCP-Industry-Grade-Implementations - LangGraph-Advanced-Workflows - n8n-Automated-Orchestrations - vLLM-Fine-Tuning-Endpoints Tech stack: Jupyter Notebook Why it matters: This repository offers a practical, step-by-step approach to building and deploying intelligent AI systems, making it invaluable for developers and engineers looking to advance their skills in AI automation and agent-based systems. Repository: MDalamin5/End-to-End-Agentic-Ai-Automation-Lab GitHub stars: 104 GitHub description: A comprehensive, hands-on engineering playbook for building, deploying, and managing intelligent AI agents, RAG pipelines, and automated workflows.

### Overview: Lollms is an all-in-one AI solution that integrates seamlessly with various AI services, providing a versatile chat interface powered by the `lollms_client` library. It features multi-user support, simplified installation, environment-based configuration, persistent discussions, integrated Retrieval-Augmented Generation (RAG), multimodal chat, advanced personality systems, and more. ### Key Features: - Multi-User Support: Secure token-based authentication. - Simplified Installation: Use `run.sh` or `run_windows.bat` scripts. - Environment-Based Configuration: Use a `.env` file for easy setup. - Persistent Discussions: Chat histories saved per user. - LLM Integration: Interact with various LLM backends using `lollms-client`. - Streaming Responses: Real-time AI responses. - Multimodal Chat: Upload images with text prompts. - Advanced Personality System: Create, edit, and delete custom personalities. - Retrieval-Augmented Generation (RAG): Organize documents into DataStores, upload files, and manage permissions. - Friend System & Direct Messaging (DM): Send, accept, and reject friend requests. - Admin Panel: Manage users, configure settings, and manage LLM/TTI/TTS bindings. - Data Export/Import: Export/import discussions and settings. - Responsive UI: Built with Vue.js and Tailwind CSS, featuring Markdown rendering, code highlighting, and math rendering with KaTeX. ### Tech Stack: - Python 3.10+ - FastAPI backend - Vue.js frontend - Tailwind CSS - `loll

Overview: This project provides an algorithm-focused interface for training, continual learning, and reinforcement learning techniques. It is designed to streamline the process of developing and optimizing models for various applications. Key features: - Training and optimization of LLMs - Continual learning capabilities - Reinforcement learning integration Tech stack: Python Why it matters: This project offers a comprehensive solution for developers and researchers looking to enhance their model training and learning processes, making it easier to implement advanced techniques in their projects.

Overview: CycleDiff is a code repository for implementing Cycle Diffusion Models for unpaired image-to-image translation, based on the TIP2026 paper. It includes a pre-trained weight and detailed instructions for training and testing the model. Key features: - Implement Cycle Diffusion Models for unpaired image-to-image translation. - Includes a pre-trained weight for efficient training. - Detailed installation and training instructions provided. Tech stack: Python Why it matters: This project is useful for researchers and developers looking to implement Cycle Diffusion Models for image-to-image translation tasks, offering a pre-trained weight for quick experimentation. Repository: ZouShilong1024/CycleDiff Primary language: Python GitHub stars: 104 GitHub description: Code for TIP2026 paper: CycleDiff: Cycle Diffusion Models for Unpaired Image-to-image Translation

### Overview: The repository contains code and models developed by the AWS AI Speech team for deep learning representations. It includes models for speech recognition, phonetic recognition, and deep contextualized acoustic representations. ### Key features: - **DeCoAR (Self-Supervised Contextual Representations for Speech Recognition)**: A model for semi-supervised speech recognition. - **BERTphone (Phonetically-Aware Acoustic BERT for Speaker and Language Recognition)**: A model for speaker and language recognition using BERT. - **DeCoAR 2.0 (Deep Contextualized Acoustic Representation with Vector Quantization)**: An improved version of DeCoAR. ### Tech stack: PyTorch and MXNet ### Why it matters: This repository provides essential tools and models for speech recognition and phonetic analysis, which are crucial in the field of artificial intelligence and speech technology.

Overview: Train in Silence is a Task-Aware MCP server designed to help developers fine-tune Large Language Models (LLMs) by providing a single, structured recommendation across multiple cloud providers. It eliminates the need for manual price comparisons, offering the cheapest, fastest, and most balanced hardware options in seconds. Key features: - Task-Aware MCP server for LLM fine-tuning - Instantly snipe the cheapest, fastest GPUs across 10+ cloud providers - Supports multiple cloud providers including Vast.ai, RunPod, AWS, and Lambda - Automatically falls back to live aggregators if API keys are not provided Tech stack: Python Why it matters: Train in Silence simplifies the process of fine-tuning LLMs by providing a comprehensive solution that eliminates manual price comparisons, saving developers time and effort. It ensures that the most efficient and cost-effective hardware options are selected automatically, making the fine-tuning process more efficient and straightforward. Repository: hlpun/Train-in-Silence Primary language: Python GitHub stars: 104 GitHub description: The first Task-Aware MCP server for LLM fine-tuning. Instantly snipe the cheapest, fastest GPUs across 10+ cloud providers.

Overview: MobileI2V is a fast and high-resolution image-to-video model designed specifically for mobile devices, achieving comparable quality with significantly reduced computational and memory requirements. Key features: - Fast inference on mobile devices - High-resolution video generation - 5.55× smaller model size compared to SVD-XT - 199× faster inference on an A100 GPU - Supports multiple video resolutions and frame rates Tech stack: Python, HuggingFace Transformers, PyTorch Why it matters: MobileI2V is crucial for applications requiring real-time video generation, such as augmented reality, interactive storytelling, and virtual reality experiences, where performance and resource efficiency are paramount.

Overview: This project provides a Python-based A2A Protocol Adapter SDK that simplifies the integration of various AI agents and frameworks into an A2A-compatible environment. It handles AgentCard generation, JSON-RPC, task lifecycle, and stream processing through the official A2A SDK, making it easier to develop and deploy AI agents without needing to write custom code. Key features: - Wraps agent CLIs, local agents, workflow engines, and Python agent frameworks. - Handles AgentCard generation, JSON-RPC, task lifecycle, and streaming. - Supports Claude Code, Codex, Pi, OpenClaw, n8n, LangGraph, LangChain, CrewAI, Ollama, Hermes Agent, and custom Python functions. - Provides two ways to use it: as a Python SDK or as a CLI. - Works with Python 3.11+ and auto-generates Agent Cards at `/.well-known/agent-card.json`. - Supports multiple frameworks and agents, including Claude Code, Codex, Pi, OpenClaw, n8n, LangGraph, LangChain, CrewAI, Ollama, Hermes Agent, and custom Python functions. - Offers a Python SDK and CLI, allowing for seamless integration and deployment of AI agents. Tech stack: Python 3.11+. Why it matters: This project streamlines the integration of AI agents into an A2A-compatible environment, making it easier to develop and deploy AI agents without needing to write custom code. It supports multiple frameworks and agents, including Claude Code, Codex, Pi,

Overview: Awesome-Text2GQL is an AI-assisted framework designed to generate high-quality Text2GQL datasets for various graph query languages, aiming to reduce the cost of annotation by automating the process. Key features: - Translation: Supports translation of graph query languages to questions in different natural languages. - Generalization: Enables question and query generalization to increase corpus diversity. - Generation: Automatically generates complex graph database schemas, data, and corpus for diverse queries. Tech stack: Python Why it matters: By automating the process of generating Text2GQL datasets, Awesome-Text2GQL reduces the need for extensive human labor, making it more efficient and cost-effective compared to traditional methods. Repository: ldbc/Text2GraphQuery-DataGen GitHub stars: 104 GitHub description: Fine-Tuning Dataset Auto-Generation for Graph Query Languages