Projects

Overview: SynthForge is a toolkit designed for generating high-quality synthetic datasets specifically tailored for large language models (LLMs) for fine-tuning. It supports multiple backends and allows for filtering, making it efficient and user-friendly. Key features: - GPU-accelerated processing - Multi-backend support - Filterable datasets Tech stack: Python Why it matters: SynthForge offers a valuable solution for developers looking to optimize the training of LLMs by providing high-quality synthetic datasets, which can significantly improve model performance and reduce the need for real-world data. Repository: Dosi-f/synthforge Primary language: Python GitHub stars: 12 GitHub description: 🧪 SynthForge — toolkit for generating high-quality synthetic datasets for LLM fine-tuning. Multi-backend, filterable, GPU-accelerated. Early-stage experimental project.

Overview: This project focuses on developing a smoke and fire detection system using custom CNN and fine-tuned Google Vision Transformer for wild environment monitoring. Key features: - Custom CNN for smoke and fire detection - Fine-tuned Google Vision Transformer for smoke and fire detection - Self-built dataset for training and evaluation Tech stack: PyTorch with PyTorch Lightning Module Why it matters: The project showcases the effectiveness of combining custom CNNs with pre-trained Vision Transformers for smoke and fire detection in forest environments, providing a robust solution for early warning systems.

### Overview: This ComfyUI-stylebook repository contains 700+ visual styles, each with a rendered preview you can browse before committing. It includes 1000+ artists with written descriptors and modifier previews for lighting, color, era, period dress, finish, and mood. The styles ship with a written description, keyword list, negative prompt, and preview image. The package is offline, zero dependencies, and no install needed. ### Key features: - **700+ Visual Styles**: Each style has a rendered preview, keyword list, negative prompt, and preview image. - **1000+ Artists**: Descriptors and written descriptions for 1000+ artists. - **Modifier Gallery**: Tilted lighting, color, era, period dress, finish, and mood modifiers with rendered tiles and verbatim text. - **No Dependencies**: No install, no downloads, no network calls, no API keys required. ### Tech stack: Python ### Why it matters: Stop guessing whether "wet plate collodion" or "risograph" will do what you want. Open the gallery, look at it, click it. Every style ships four things: a written description, a keyword list, a matching negative prompt, and a preview image. No dependencies, no downloads, no network calls, no API keys. Every style ships four things: a written description, a keyword list, a matching negative prompt, and a preview image. 1050+ artists ship a written descriptor each, so the look still

Overview: This project is an application that uses Weaviate, a vector-based image similarity search engine, to find similar products across the market. The app leverages 62k images of products combined with their text descriptions to build vectors for search purposes. Key features: - Vector-based image similarity search for product recommendations - Integration with Weaviate for efficient image and text data storage and retrieval - Use of 62k product images and descriptions for accurate search results Tech stack: Python Why it matters: This project is useful for stores to identify if their products were also in their competitor's offerings, providing valuable insights into market trends and competition.

Overview: TechRAG is a hybrid ontology-driven retrieval engine designed to enhance the precision and traceability of AI-generated technical documentation. Key features: - Vector embeddings via LanceDB (OpenAI-compatible endpoint) - Domain ontology + knowledge graph following the SQLGraph approach - Atomic facts with source back-reference - Response audit metadata (evidence coverage, reasoning score, recovery cycles) Tech stack: Python Why it matters: TechRAG leverages a structured knowledge representation and adaptive vector search to provide precise and traceable context to LLMs, ensuring better quality and reliability of technical documentation. Repository: FBR65/TechRAG GitHub stars: 12 GitHub description: TechRAG is a hybrid ontology-driven retrieval engine that combines a SQLAlchemy-based knowledge graph and LanceDB vector store with OpenAI-compatible chat and embedding clients for retrieval-augmented generation with audit metadata and recovery cycles.

Overview: COVID-VTS is a fact-checking dataset designed for short video platforms, focusing on automatically generating verifiable, trustworthy, and misleading claims on the platform. Key features: - Automatic generation of large-scale verifiable, trustworthy, and misleading claims - Proposed TwtrDetective framework for fact-checking on short video platforms Tech stack: Python, Golang, TensorFlow Why it matters: The dataset and framework provide a practical tool for researchers and practitioners to evaluate and improve fact-checking algorithms on short video platforms, contributing to the field of digital misinformation detection.

Overview: This repository contains open-source skills designed to maintain brand consistency in AI-generated visuals, specifically for Claude. It addresses the 30% problem of style drift by providing a disciplined, training-free workflow that ensures visual identity remains consistent across generations. Key features: - Style-consistent image & video generation - Disciplined, training-free workflow for brand consistency - Accurate, corrective guidelines for specific failures - Representative atlas for reference-based editing Tech stack: Python, PyTorch, and other dependencies. Why it matters: This approach offers a practical solution for maintaining brand consistency in AI-generated visuals, avoiding the downsides of LoRA training and ensuring high-quality, brand-consistent outputs without the need for expensive retraining.

Overview: This project aims to provide a demo of QDrant, a vector database, for NLP tasks, focusing on interactive query creation and display of results. Key features: - Supports creating and storing queries for QDrant Vector Database. - Displays positives, negatives, and results for each query. - Enables downloading datasets directly from Hugging Face. - Provides a loading sign while generating embeddings. - SQLiteDB for storing query results and names. - SQLiteDB for storing query results and names. Tech stack: Python Why it matters: This project serves as an introduction to vector databases, offering a simple way to interact with QDrant for NLP tasks. It demonstrates the capabilities of QDrant and provides a foundation for more advanced use cases. Repository: GeorgePearse/QDrant-NLP Primary language: Python GitHub stars: 12 GitHub description: QDrant-NLP

Overview: This repository showcases the implementation of a basic RAG (Reinforcement Augmented Generation) pipeline using LangChain in Supabase Edge Functions, providing a practical example for developers. Key features: - Basic RAG pipeline using LangChain - Integration of LangChain packages in Supabase Edge Functions - Installation of npm packages via `import_map.json` - Deployment of Supabase Edge Functions to production Tech stack: TypeScript Why it matters: This project offers a clear demonstration of how to integrate LangChain with Supabase Edge Functions, making it a valuable resource for developers looking to enhance their applications with advanced natural language processing capabilities.

### Overview: VelociRAG is a lightning-fast RAG solution designed for AI agents, leveraging ONNX Runtime for efficient vector search and four-layer retrieval fusion. It provides sub-200ms warm search times and integrates seamlessly with MCP servers, making it ideal for real-time applications without the need for GPUs or PyTorch. ### Key Features: - **4-layer Retrieval Fusion:** Combines vector similarity, BM25 keyword matching, knowledge graph traversal, and metadata filtering. - **ONNX Runtime:** Runs all components on local models, ensuring no GPU or PyTorch dependencies. - **Incremental Indexing:** Adds documents without rebuilding the entire index, making it easier to manage large datasets. - **MCP Server Integration:** Plugs into Claude, Cursor, and Windsurf for agent integration, providing a Unix socket daemon for warm queries. - **CLI and Python API:** Available for both command-line and Python integration, allowing seamless integration with various AI agents. ### Tech Stack: VelociRAG is built using Python, utilizing ONNX Runtime for inference and FAISS for vector similarity search. The solution is designed to be lightweight and efficient, making it suitable for real-time applications. ### Why It Matters: VelociRAG offers a powerful and efficient way to integrate AI agents into existing systems, providing fast, accurate search capabilities without the need for complex frameworks like PyTorch. Its lightweight nature and easy integration with MCP servers make it a valuable tool for developers looking to enhance their AI applications.

Overview: AgriMind is a smart application designed to assist farmers in various agricultural tasks, providing accurate and straightforward answers to their questions. It is built on a combination of LLMs and micro-tuned models to address specific agricultural issues. Key features: - Crop disease diagnosis and treatment recommendations - Fertilizer suggestions and planting techniques - Veterinary disease diagnosis and feed recommendations - Animal health management and care - Agricultural equipment and machinery troubleshooting and maintenance Tech stack: Python, LLMs, and micro-tuned models. Why it matters: AgriMind simplifies complex agricultural problems for farmers, ensuring they can make informed decisions with minimal input and guidance from the model. Repository: JimmylearningCoding/AgriMind Primary language: Python GitHub stars: 12 GitHub description: AgriMind: The smart LLM application that every farmer truly needs. README excerpt: <!-- Improved compatibility of back to top link: See: https://github.com/othneildrew/Best-README-Template/pull/73 --> <br /> <div align="center"> <a href="https://github.com/JimmylearningCoding/AgriMind"> <img src="images/wukong.jpg" alt="Logo" width="130" height="130"> </a> <h1 align="center">AgriMind</h1> <p align="center"> AgriMind: The smart ally that every farmer truly needs. <br /> <a href="https://github.com

### Overview: CatIss is an intelligent tool designed for automatic categorization of issue reports based on the RoBERTa model. It has been recognized as the winner of the NLBSE'22 tool competition, achieving an impressive F1-score of 87.2% on the provided test set. ### Key Features: - **Automatic Categorization**: Categorizes issue reports into three main categories: bug report, enhancement/feature request, and question. - **Fine-Tuning RoBERTa**: Fine-tuned on a preprocessed dataset from the NLBSE 2022 Tool Competition. - **Wide Application**: Generic model applicable to any unseen software project or projects with limited historical data. ### Tech Stack: - **Model**: RoBERTa - **Datasets**: Cleaned and preprocessed datasets from the NLBSE 2022 Tool Competition. ### Why It Matters: CatIss is a significant advancement in issue report categorization, offering high accuracy and broad applicability across various software projects. Its winning performance in the NLBSE'22 competition underscores its effectiveness and reliability in real-world scenarios. ### Repository: - **Repository**: MalihehIzadi/catiss - **Primary Language**: Jupyter Notebook - **GitHub Stars**: 12 ### Shared Model and Data: - **Model and Datasets**: Saved model and cleaned datasets are available on Google Drive. ### System Information: - **Hardware**: Ubuntu 16.04, two GeForce RTX 2080 GPU cards, AMD Ryzen Thread

### Overview: QuantumLeap is a high-performance language model framework that leverages the expertise of llama.cpp to achieve 130% faster MoE inference. Built on the latest advancements in expert flow and kv compression, it optimizes the performance of large language models, making them more efficient and faster to run on various hardware platforms. ### Key Features: - **TurboQuant KV Compression**: Dramatically reduces the size of the model by up to 7.4x, significantly reducing memory usage. - **ExpertFlow MoE Optimization**: Enhances model performance by up to 2x, thanks to intelligent caching and adaptive prefetching. - **Ollama API Compatibility**: Runs on its own port, ensuring seamless integration with Ollama without the need to uninstall llama.cpp. - **130% Faster Inference**: Achieves 4.34 tok/s on a 6GB VRAM setup, compared to 1.89 tok/s in Phase 2, with a 2.3x performance improvement. - **Real-World Results**: Tested on a RX 5600 XT 6GB GPU, achieving 4.34 tok/s, showcasing a 130% performance boost. ### Tech Stack: Built on llama.cpp, QuantumLeap is designed for maximum performance and efficiency. ### Why It Matters: QuantumLeap simplifies the process of running large language models by providing a production-ready framework that optimizes performance and reduces resource usage. It is designed to be easily integrated into existing systems, making it ideal for

Overview: Qwen-Image-2.1 is a text-to-image model that renders prompts and photos on a 12 GB NVIDIA GPU in about 35 seconds, making it ideal for users with limited GPU resources. This model is designed to be small and efficient, with a 7B generator, making it suitable for those without a 24 GB card. Key features: - Renders prompts and photos on a 12 GB NVIDIA GPU - Generates images in about 35 seconds per 1024² image, using an RTX 3060 - Supports up to 10 reference photos per edit - Runs on a desktop app and CLI - Benchmarked against GPT-Image-2, showing significant performance improvements Tech stack: Rust Why it matters: Qwen-Image-2.1 offers a powerful image generation solution for users with limited GPU resources, making it accessible to a broader audience. Its small model size and efficient rendering capabilities make it a valuable tool for developers and content creators looking to enhance their projects without the need for a high-end GPU.

### Overview: NeuralCache is a lightweight reranker for Retrieval-Augmented Generation (RAG) pipelines that remembers which context the model actually uses. It blends dense semantic similarity with a narrative memory of past wins and stigmergic pheromones to optimize relevance without rebuilding the stack. ### Key Features: - **Narrative Memory (EMA)**: Keeps track of passages that consistently helped users, biasing future reranks toward them. - **Stigmergic Pheromones**: Reward useful documents but decay over time, preventing filter bubbles. - **MMR + ε-greedy**: Introduces diversity without tanking relevance. - **Zero External Dependencies**: Uses a hashing trick for embeddings so you can see results instantly. ### Tech Stack: Python ### Why It Matters: NeuralCache is an opinionated, stateful reranking layer designed to increase practical usefulness of RAG retrieval results by remembering what historically mattered, decaying stale signals, maintaining diversity, and optimizing compute via intelligent gating. The repository is production-minded (CI, packaging, adapters, metrics) yet approachable with minimal dependencies out of the box. Its architecture cleanly separates scoring components, adapters, and API surfaces, making it a solid foundation for iterative improvement and integration into existing LLM pipelines. --- Repository: **Maverick0351a/neuralcache** Primary Language: **Python** GitHub Stars: **12** --- ### ⚡ 60-second Quickstart ```bash # 1. Install pip install neuralcache # 2. Launch

Overview: RAG-nificent is a production-ready Chainlit application that leverages Retrieval-Augmented Generation (RAG) to provide instant answers and references from a curated directory of PDFs. It supports the latest GPT-5, GPT-4.1, and advanced LLM models, making it a powerful tool for researchers, policy makers, and the public seeking specific information from extensive documents. Key features: - **Conversational Interface**: Engage with the system using natural language queries to receive responses directly sourced from the PDFs. - **Direct Citation**: Every response includes a direct link to the source PDF page, ensuring traceability and verification. - **PDF Directory**: A predefined set of key PDF documents, including WHO recommendations on major health topics. Tech stack: Python, LangChain, Chainlit, Pinecone, OpenAI, Groq Models, and LLMs. Why it matters: RAG-nificent offers a user-friendly interface for researchers and policymakers to quickly find specific information within extensive documents, enhancing efficiency and accuracy in their work.

Overview: This repository contains code for training landcover classification models on Sentinel-2 data using EfficientNet-Unet and Prithvi, with labels from OpenStreetMap and CNES. Key features: - EfficientNet-Unet model for landcover classification - Prithvi foundation model for fine-tuning - Labels from OpenStreetMap and CNES - Data download and visualization tools Tech stack: Python Why it matters: This project provides a comprehensive solution for landcover classification using Sentinel-2 data, leveraging advanced models and high-quality labels for improved accuracy and reliability. GitHub stars: 12 Repository: MaxWolf-01/sentinel2-landcover-classification

Overview: Metatron2 is an easy-to-use Discord bot that leverages machine learning functions, including LLM chat, image generation, and speech generation capabilities. It supports local or API-based generation and provides features like multimodal LLM chat, Stable Diffusion image generation, Bark audio generation, and voice cloning. Key features: - Multimodal LLM chat - Stable Diffusion image generation - Stable Diffusion XL image generation - Bark audio generation and voice cloning - Support for local generation or using APIs for generation Tech stack: Python Why it matters: Metatron2 is a versatile Discord bot that integrates multiple cutting-edge AI and machine learning functionalities, making it a valuable tool for developers and enthusiasts interested in creating interactive and dynamic chatbots.

Overview: This project provides a unified CPU wheel optimized for different instruction sets, enabling efficient inference of large language models (LLMs) without the need for a GPU. Key features: - Unified CPU wheel with automatic ISA detection - Supports AVX2, AVX-512, VNNI, BF16, AMX, NEON, FP16, DOTPROD - Compatible with PyPI and Docker Tech stack: Shell Why it matters: This project simplifies the process of running LLMs on CPUs, making it accessible to a wider audience and reducing the reliance on expensive GPU hardware.

### Overview: Comfy工作台是一款原生Android应用，专为ComfyUI用户提供定制化、便捷的图形化编辑体验。内置6个内置模板，无需网络连接即可出图。支持多种尺寸和参考视频，提供简洁的参数输入界面。 ### Key features: - **内置模板**：6个预设工作流，一键生成高质量图片。 - **参数输入**：提示词、尺寸、种子、模型、参考图等参数通过输入框直接设置。 - **实时反馈**：生成过程中的进度和错误信息实时显示。 - **断线续传**：提交前落库，断线后任务以队列和历史为准，重连后继续处理。 - **LoRA与参考视频**：支持LoRA加载器和参考视频导入。 - **提示词库**：提供分类取词、随机抽卡、保存和编辑功能。 - **画廊**：统一浏览图片和视频，支持放大和播放，不自动写入系统相册。 - **后台运行**：生成和下载任务在前台服务中进行，后台运行稳定。 - **更新检查**：每天自动检查更新，支持Gitee和GitHub两种方式。 ### Tech stack: Kotlin + Jetpack Compose ### Why it matters: Comfy工作台简化了ComfyUI的使用流程，提供了直观的图形化编辑体验，使得用户可以随时随地进行图片生成、反推提示词等操作。通过内置模板和参数输入界面，用户可以快速生成高质量