Projects

Overview: CodeGraph turns any codebase into a Neo4j knowledge graph, layering vector embeddings for semantic search, enabling natural-language queries that traverse structure and rank by meaning. Key features: - GraphRAG hybrid graph traversal combined with vector search - One queryable database with knowledge graph, vector index, and file metadata in Neo4j - Natural-language queries in plain English, grounded in real code - Streamlit interface with live connection status, workflow tracking, and real-time feedback - Configurable models for OpenAI entity extraction and model tuning - One-command Docker for easy setup and configuration Tech stack: Python, Neo4j, OpenAI embeddings, Streamlit, Docker Why it matters: CodeGraph bridges structure and meaning, providing a single, queryable database that combines graph traversal and semantic search, making it a powerful tool for understanding and querying any codebase.

Overview: RAPTOR is an open-source multimodal AI framework and agent harness designed for building agentic applications using video, audio, images, documents, and organizational knowledge. It includes MCP tools, A2A orchestration, persistent memory, hybrid RAG, GraphRAG, evaluation, guardrails, and model lifecycle services. Key features: - MCP Server - Memory Service - Personal DB Service - Guardrail Service - Benchmark Service - Video Search 2.0, GraphRAG, A2A, branch isolation Tech stack: Python Why it matters: RAPTOR provides a comprehensive solution for enterprise AI applications, enhancing search, memory management, and user isolation, making it a valuable tool for developers looking to build robust, scalable, and secure AI-driven applications.

### Overview: This repository provides a comprehensive guide on building sovereign RAG systems using Oracle AI Database 26ai, focusing on advanced methodologies such as Dual RAG, Graph RAG, and multimodal video pipelines. It demonstrates how to integrate AI capabilities directly into the database, ensuring data security and governance. ### Key Features: - **Dual RAG**: Synchronizes unstructured vector semantics with deterministic structured SQL to combat hallucinations and data poisoning. - **MAS-RAG (Multi-Agent Systems for RAG)**: A Universal Context Engine with a Planner, Agent Registry, and sovereign Oracle agents that replace external vector databases. - **Hybrid Adaptive RAG**: A dynamic inference-time router that switches between model reasoning, RAG retrieval, and expert human feedback injection based on cosine similarity scores and user rankings. - **Converged Hyper-Query**: Fuses vector similarity, Oracle Spatial (SDO_GEOMETRY), and SQL Property Graph (SQL/PGQ) traversal in a single execution context without data movement. ### Tech Stack: Jupyter Notebook ### Why It Matters: This repository offers a practical approach to integrating AI capabilities directly into Oracle's converged database engine, Oracle AI Database 26ai, thereby eliminating the fragmentation, latency, and security risks associated with traditional data extraction.

Overview: MnMS is an agent-based multimodal urban mobility simulator designed for large-scale urban environments, resulting from the ERC MAGnUM project. It simulates multimodal traffic and considers various mobility services, including traditional and new mobility options. Key features: - Simulates multimodal dynamic traffic for urban environments. - Models travel choices considering all multimodal options. - Governs vehicle motion using regional multimodal MFD curves. - Addresses large-scale research topics in urban mobility management, regulation, and optimization. Tech stack: Python Why it matters: MnMS is a powerful tool for urban mobility researchers and practitioners, enabling the simulation and analysis of complex urban transportation systems. Repository: EMob-Lab/MnMS GitHub stars: 22 GitHub description: Agent-based Multimodal Urban Moblity Simulator resulting from the ERC MAGnUM project

Overview: This repository contains a guide site for skill-based text-to-image prompt engineering, focusing on three specific prompt types: figure-photo, youth-seduction, and erotic. The site includes a director's table, writing rules, and a combination gallery. It is designed for 18+ users and features explicit content, with a strict age restriction policy. Key features: - Skill-based prompt engineering guide - Three types of prompts: figure-photo, youth-seduction, and erotic - Director's table for skill combination assembly - Writing rules for skill usage - Combination gallery for skill-specific combinations - Public dimension system with L1-L4 levels and dimension categories - Public dimension schema and point-and-dish strategy - Public dimension schema and point-and-dish strategy - Public dimension schema and point-and-dish strategy - Public dimension schema and point-and-dish strategy - Public dimension schema and point-and-dish strategy - Public dimension schema and point-and-dish strategy - Public dimension schema and point-and-dish strategy - Public dimension schema and point-and-dish strategy - Public dimension schema and point-and-dish strategy - Public dimension schema and point-and-dish strategy - Public dimension schema and point-and-dish strategy - Public dimension schema and point-and-dish strategy - Public dimension schema and point-and-dish strategy - Public dimension schema and point-and-dish strategy - Public dimension schema and point-and-dish strategy - Public dimension schema and point-and-dish strategy - Public dimension schema and point-and

Overview: `claude-db` is a multi-paradigm database design, audit, and migration toolkit for Claude Code, designed to help developers ensure their schema is both structurally sound and performant, with the option to safely migrate data. Key features: - **Multi-paradigm Support**: Supports SQL, document, key-value, vector, time-series, and graph databases. - **Two Score Scores**: Measures Design & Integrity and Performance & Scale separately, providing clear insights into schema quality and performance. - **Offline Mode**: Works fully offline, including Tier 0 live introspection. Tech stack: JavaScript Why it matters: `claude-db` offers a comprehensive solution for database design and management, ensuring that developers can confidently model, audit, and migrate their Claude Code schemas without sacrificing performance or data integrity. Repository: Hainrixz/claude-db GitHub stars: 22 GitHub description: The multi-paradigm database expert for Claude Code — design a schema, audit it on two never-blended scores (Design & Integrity / Performance & Scale), and safely migrate. SQL · document · key-value · vector · time-series · graph. Works offline. Why: Most database tooling focuses on style checking, query execution, or ORM abstraction. `claude-db` provides a holistic view of schema design, performance, and migration, ensuring that developers can make informed decisions about their database infrastructure.

### Overview: The BoDmagh dataset is a Supervised Fine-Tuning (SFT) dataset for the Darija language, meticulously curated by ImadSaddik. It includes conversations between a user and an assistant, structured in JSON format and available on Hugging Face. ### Key features: - Structured JSON format - Conversations between a user and an assistant - Includes metadata fields such as `id`, `conversation`, `token_count`, `turns_count`, `markdown_conversation`, `source`, `topic`, and `safety_flags` - Updated daily to ensure high quality ### Tech stack: Jupyter Notebook ### Why it matters: The dataset is essential for fine-tuning models in the Darija language, providing high-quality conversations for training and evaluation purposes.

### Overview: Artist Image Generator: DALL·E AI-Driven Image & Avatar Creator for WordPress Developed by Pierre Viéville for the WordPress community, Artist Image Generator is an AI-driven image and avatar creator plugin powered by DALL·E 2 and DALL·E 3. Seamlessly integrate OpenAI's state-of-the-art technologies into popular page builders, including Gutenberg, Elementor, Beaver Builder, WP Bakery, and Woocommerce. ### Key Features: - **Image Generations**: Create original images based on text prompts. - **Image Variations**: Generate diverse image variations effortlessly. - **Image Edits**: Unlock advanced editing features by uploading a mask. - **Shortcodes**: Prompt form to generate and download images. - **WordPress Avatar AI**: Create AI-generated avatar for WordPress users. ### Tech Stack: PHP ### Why It Matters: Artist Image Generator revolutionizes image creation, offering seamless integration with popular WordPress page builders and the power of DALL·E 2 and DALL·E 3 for AI-driven image generation and editing. ### Installation: - **Search and Install**: Search for the Artist Image Generator plugin in the WordPress extensions library and click "Install." - **Download and Upload**: Alternatively, download the zip archive file and upload it through the "upload plugin" tools. - **Activate Plugin**: Click "Activate" and the plugin is ready to configure. ### Configuration: - **OpenAI API Key**: Sign up / Log in to the OpenAI developer portal and generate an API key. ### GitHub Stars: 22 ###

Overview: softrag is a minimal, local-first multimodal RAG library powered by SQLite, designed for easy deployment and portability. Key features: - Supports vector and full-text indexing in a single SQLite file. - Embeds models from the environment or uses a dependency-free hash embedder. - Works fully offline, without requiring a server or vendor account. - Provides metadata filtering and document extraction capabilities. Tech stack: Python Why it matters: softrag simplifies the development of RAG systems by eliminating the need for a server and vendor accounts, making it ideal for developers looking to quickly prototype and deploy their own RAG solutions.

Overview: This project offers a suite of command-line interface (CLI) tools designed to streamline AI tasks, including speech, text, and language processing. Key features: - Speech recognition - Text summarization - Language translation Tech stack: Python Why it matters: These tools simplify complex AI tasks, making them accessible and easy to integrate into various applications and projects.

### Overview: FigureWeave is a research-engineering project designed to convert method descriptions into editable scientific figures. It leverages local GPU SAM3 segmentation for efficient processing and supports dual-provider routing for enhanced flexibility. The project aims to generate publication-quality figures that can be easily modified and reused in various formats, such as method overviews, pipeline diagrams, system schematics, and architecture figures. ### Key Features: - **Local GPU SAM3 Segmentation**: Efficiently segments images using local GPU SAM3 for faster processing. - **Dual-Provider Routing**: Supports image drafting and SVG reasoning with multiple providers for enhanced flexibility. - **Multi-Candidate End-to-End Generation**: Generates multiple candidate figures for improved accuracy and diversity. - **Figure Caption Conditioning**: Adds caption conditioning alongside method text for better figure context. - **SVG First Reconstruction**: Reconstructs figures using SVG-first techniques, optimizing template, template, and final assembly stages. - **CUDA-Accelerated Post-Processing**: Post-processing is optimized with CUDA-accelerated local post-processing for segmentation and background removal. ### Tech Stack: - **Python 3.10+**: The primary programming language used for the project. ### Why It Matters: FigureWeave is particularly useful for method overviews, pipeline diagrams, system schematics, and architecture figures. It allows for the creation of editable draft figures that can be easily modified and reused across various formats, such as papers, slides, and reports. The project is not intended to replace precise plotting tools like matplotlib, seaborn, ggplot, or Origin, as it focuses

Overview: PLMFit is a platform designed to enable researchers with varying levels of computational expertise to fine-tune Protein Language Models (PLMs) using simple command-line instructions. It democratizes the use of deep learning for researchers in the field of laboratory research. Key features: - Command-line interface for easy execution - Support for multiple PLMs including ESM, ProGen2, and ProteinBERT - Data management and configuration through environment variables - SLURM scripts optimized for the ETH Zurich supercomputing cluster Tech stack: Python Why it matters: PLMFit simplifies the process of fine-tuning PLMs, making advanced research accessible to a broader audience without requiring extensive programming knowledge. It also optimizes computational resources, particularly for researchers working with large datasets or on high-performance computing clusters. Repository: LSSI-ETH/plmfit Primary language: Python GitHub stars: 22 GitHub description: PLMFit platform for TL on PLMs

### Overview: IGenBench is a benchmark designed to evaluate the reliability of text-to-infographic generation. It includes 600 test cases across 10 reliability dimensions, aiming to assess whether the generated infographics are correct, complete, and faithful to the prompt. ### Key features: - **600 Test Cases**: Covers 10 reliability dimensions. - **10 Dimensions**: Data completeness, encoding, ordering, accuracy, completeness, ordering, completeness, accuracy, and order. - **Model Evaluation**: Scores the model's performance on each dimension. ### Tech stack: Python ### Why it matters: IGenBench provides a comprehensive benchmark to evaluate the reliability of text-to-infographic generation models, ensuring that the generated infographics are accurate, complete, and faithful to the input prompt. ### Repository: MisterBrookT/IGenBench ### GitHub stars: 22 ### Primary language: Python ### GitHub description: [ACL 2026] A benchmark for evaluating the reliability of text-to-infographic generation with curated test cases and automated question-based verification.

### Overview Tomechat is a chatbot application built on the LangChain framework, utilizing FAISS vector search for knowledge-base retrieval. It connects to remote LLM services via OpenAI-compatible APIs or the Google Gemini SDK, implementing knowledge-base RAG chat with FAISS vector search. The backend is powered by FastAPI, while the frontend is built with Next.js, and SQLite and FAISS are used for knowledge-base metadata storage and vector retrieval. ### Key Features - **LLM Chat**: Accesses remote LLMs through OpenAI-compatible APIs or the Google Gemini SDK. - **Knowledge-Bas RAG**: Implements local knowledge-base Q&A using FAISS vector retrieval. - **Search Engine Chat**: Integrates Web search capabilities via Tavily (or self-hosted SearxNG) for enhanced chat functionality. - **Knowledge-Bas Management**: CRUD operations, document vectorization, and FAISS index management. ### Tech Stack - **LangChain Framework**: The backend uses FastAPI for the backend and Next.js for the frontend. - **FAISS Vector Search**: Used for knowledge-base vector retrieval. - **SQLite + FAISS**: Stores and retrieves knowledge-base metadata. - **OpenAI/Google Gemini SDKs**: Connects to remote LLM services. - **Node.js 20+**: Used for the frontend. ### Why It Matters Tomechat is designed to be a versatile chatbot capable of integrating with multiple LLM services, providing a robust knowledge-base RAG chat experience, and enhancing the search capabilities through external search engines. The Lang

### Overview: Holmes is an agent skill that reads photos like Sherlock Holmes, observing first and deducing second. It works with Claude Code, Codex, Cursor, Gemini CLI, OpenCode, and GitHub Copilot. ### Key Features: - **Observation Sealing:** Everything the agent sees is listed and sealed before reasoning begins. - **Evidence Citing:** Each conclusion cites the observations it is based on, offering a chain of at most three steps and a calibrated confidence. - **Grading:** You mark each conclusion as a hit, half, or miss, and flag any observations not present in the photo. - **Commitment:** At least one conclusion must combine to form a valid conclusion. ### Tech Stack: TypeScript ### Why It Matters: Holmes is designed to be a reliable and efficient tool for analyzing photos, ensuring that the conclusions are based on solid evidence and can be easily graded by the user. It is particularly useful for tasks that require a clear and logical reasoning process, such as identifying patterns or making deductions based on observations.

Overview: This repository is a benchmark framework designed to evaluate and compare different LLM compression algorithms, focusing on their performance in tasks like MMLU, ARC, and perplexity measurement. Key features: - Supports multiple compression algorithms: GPTQ and picoLLM Compression - Meets the needs of developers and researchers in the field of language model compression - Provides a minimalistic and extensible framework Tech stack: Python Why it matters: This benchmark is crucial for developers and researchers in the field of language model compression, as it allows for a standardized comparison of different compression techniques, aiding in the selection of the most effective method for specific use cases. Repository: Picovoice/llm-compression-benchmark Primary language: Python GitHub stars: 22 GitHub description: LLM Compression Benchmark

Overview: This repository provides an easy-to-deploy web playground for the Stable Diffusion model, allowing users to access and experiment with the model without the limitations of existing platforms. Key features: - Deployable web playground for Stable Diffusion model - Customizable options for the diffusion model - Easy deployment via Huggingface web demo - Full access to the original FP32 weights for precise results Tech stack: Python Why it matters: This project offers a straightforward solution for users to explore and utilize the Stable Diffusion model, providing a more flexible and accessible environment compared to existing platforms. Repository: Raibows/WebPlayground-of-StableDiffusion GitHub stars: 22 GitHub description: An Easy Deployment Web Playground of Stable Diffusion Model

Overview: Pyrag-Core is an open-source Python-first Retrieval-Augmented Generation (RAG) platform built on FastAPI APIs and LangGraph-native agents, designed for efficient and scalable retrieval-based generation tasks. Key features: - Python-first architecture - FastAPI APIs for API-driven development - LangGraph-native agents for seamless integration - Pluggable vector stores (Qdrant, Weaviate, Milvus, pgvector, Elasticsearch) - Multi-provider LLM support (OpenAI, Anthropic, Gemini, Ollama, vLLM) Tech stack: Python 3.12 + Docker Why it matters: Pyrag-Core offers a robust and scalable solution for developers looking to build production-ready RAG platforms, leveraging the strengths of Python and LangGraph for efficient retrieval-based generation tasks. Repository: SheikhSarvar/pyrag-core GitHub stars: 22 GitHub description: Open-source, Python-first RAG platform with FastAPI APIs, LangGraph-native agents, pluggable vector stores (Qdrant/Weaviate/Milvus/pgvector/Elasticsearch), and multi-provider LLM support (OpenAI/Anthropic/Gemini/Ollama/vLLM).

### Overview: CV Maker Engine is an open-source, local-first CV & resume management platform that offers privacy-focused semantic search. It instantly matches your skills, projects, and experiences to job mandates, allowing you to track your application pipeline and tailor resumes locally without subscription fees or cloud privacy concerns. ### Key Features: - **Application Dashboard & Kanban:** Track your application pipeline stage-by-stage with an interactive Kanban board. - **Browser Extension Extractor:** Extract job offers directly from job boards using the companion browser extension. - **Analytics & Keyword Insights:** Real-time stats on your job hunt, status breakdowns, and key tech stack/keyword frequency extraction. - **Hybrid Matching Engine:** Lightweight local NLP/RAKE for keyword matching and technology extraction, vector & RAG search for semantic similarity scores, and an automated flexible AI pipeline. - **Tailored Resume Builder:** Dynamically show/hide sections or sub-blocks, match bullet points to mandates, and export clean PDFs. - **Local-First Architecture:** Your job search and master profile stay saved locally on your machine. ### Tech Stack: - Frontend: React, TypeScript, Tailwind CSS, shadcn/ui. - State Management: Zustand (Global store). - Runtime: Electron (Secure IPC communication via contextBridge). - Database: SQLite (better-sqlite3) for metadata persistence and local vector storage. - AI & NLP: - Embeddings: transformers.js (Model: Xenova/all-MiniLM-L6-v2) for local vectorization. - **Local LLM

### Overview: This repository contains the implementation of a novel optimization technique called FSAM (Fisher Sharpness-Aware Minimization) for language models, specifically designed for better generalization on language models. The method aims to improve the performance of language models by leveraging the Fisher information matrix to guide the optimization process more effectively. ### Key Features: - **FSAM Optimizer**: Implements the Fisher Sharpness-Aware Minimization (FSAM) technique to enhance the generalization of language models. - **PyTorch Backend**: The implementation is written in PyTorch, ensuring compatibility and ease of use. - **Fine-Tuning Scripts**: Provides detailed scripts for fine-tuning language models like RoBERTa, including training parameters and examples. ### Tech Stack: - Python 3.8+ - PyTorch >= 1.10.0 - NVIDIA GPU for training - NCCL for optimal performance ### Why It Matters: - **Improved Generalization**: FSAM optimizes the training process to improve the generalization of language models, leading to better performance on downstream tasks. - **Versatile Application**: The technique can be applied to various language models, enhancing their ability to generalize across different tasks. ### GitHub Stars: - 22 stars indicate that the repository has gained some traction and interest in the community. ### Repository: - Repository: [WHU-ZQH/FSAM4PLM](https://github.com/WHU-ZQH/FSAM4PLM) - Primary Language: Python - GitHub Stars: 2