Projects

Overview: LumeAI is an AI-powered multimodal chat application built with Next.js, featuring advanced features like multiple AI model support, advanced memory system, real-time conversation title generation, and customizable settings. Key features: - Seamless integration with multiple AI providers (OpenAI, Anthropic, OpenRouter) - Advanced memory system for contextual conversations - Dynamic system prompts and conversation settings - Web search capabilities with Google Custom Search integration - Web page content fetching and processing - Real-time conversation title generation - Adjustable context length and token limits - Temperature and Top-P parameter controls - Secure API key encryption - Modern, responsive interface - Dark mode support Tech stack: TypeScript Why it matters: LumeAI offers a cutting-edge chat application with advanced AI integration and customizable features, making it ideal for developers looking for a robust and user-friendly chat platform. Repository: C-W-D-Harshit/lume-ai Primary language: TypeScript GitHub stars: 15 GitHub description: AI-powered multimodal chat app with real-time responses, file support, token tracking, and dark mode. Built with Next.js. Open source under MIT.

### Overview: Serve LLMs on NCSA hardware. Supports the best FOSS models and the long tail on HuggingFace Hub. ### Key features: - **Drop-in Replacement for OpenAI Calls**: Perfectly replicates the OpenAI API, making it a drop-in replacement for OpenAI calls (except embeddings). - **Ultra-Low Latency Inference**: Runs on NVIDIA GPUs, providing ultra-low latency and high performance. - **Support for 100% of HuggingFace Hub Models**: Supports all popular models on the HuggingFace Hub, with optimized inference CUDA kernels. - **Customizable and Fused CUDA Kernels**: Suitable for building other applications on top of the model. - **Persistent and Ultra-Low Latency Inference**: Provides persistent, ultra-low latency inference with customized, fused, CUDA kernels. - **Custom Pre/Post-Processing**: Supports models requiring custom pre/post-processing code, with documentation to come. ### Tech stack: Python ### Why it matters: It's an API: I host it, you use it. Quick and easy for jobs big and small. Access it however you like: Python client, Curl/Postman, or a full web interface playground. Running on NCSA Center of AI Innovation GPUs, fully private & secure thanks to https connections via Zero Trust CloudFlare Tunnels. Supports the best FOSS models and the long tail on HuggingFace Hub.

Overview: GPT-4 eyes have been given to a vision model, allowing it to accurately identify objects and search the web for additional information. Key features: - GPT-4 eyes - TensorFlow Models - MobileNet - GPT API - Object identification - Web search Tech stack: JavaScript Why it matters: This project showcases the integration of AI with traditional computer vision, demonstrating how GPT-4 can enhance the capabilities of a vision model, providing accurate object identification and supplementary information. Repository: Charmve/gpt-eyes GitHub stars: 15 GitHub description: I GAVE GPT-4 EYES!

Overview: This agent skill prepares a Linux server with NVIDIA or Intel Arc GPUs for neural-network and language model (LLM) workloads. It includes a verified Docker + GPU inference stack, driver setup, kernel headers verification, Intel Arc driver, Docker passthrough, staged verification gates, GPU power limits, and cooling solutions. Key features: - NVIDIA driver and CUDA - Mandatory kernel headers - Intel Arc driver (Alchemist and Battlemage) - Docker + NVIDIA Container Toolkit - Staged verification gates - Inference presets (optional) - GPU power limit - Intel Arc cooling and idle power fixes Tech stack: Python Why it matters: This skill provides a reliable and verified setup for running large language models and neural networks on Linux servers, ensuring optimal performance and stability. Repository: EvilFreelancer/gpu-server-setup Primary language: Python GitHub stars: 15 GitHub description: Agent skill that prepares a Linux server with NVIDIA or Intel Arc GPUs for LLM workloads — driver, Docker with verified GPU access, staged verification gates, compose presets for vLLM, llama.cpp, Ollama, OpenWebUI and OpenVINO, persistent power caps, GPU-driven fan control and Intel Arc idle-power fixes.

Overview: This project introduces a multi-modal fusion method using VGG16 and FastText to classify information from social media platforms. Key features: - Utilizes VGG16 and FastText for multimodal data fusion - Integrates images and text messages for improved classification accuracy - Provides a dataset of images and text messages for training and testing Tech stack: Python Why it matters: This approach offers a novel method for identifying useful information from social media, potentially enhancing the effectiveness of social media monitoring and analysis.

**Overview:** MM-FM is a multimodal flow matching framework for generating high-quality images and videos. It achieves state-of-the-art performance in the CVPR 2026 competition on multimodal distribution flow matching. **Key features:** - **Flow Matching for Multimodal Distributions**: MM-FM uses flow matching techniques to align and synthesize multimodal distributions. - **Mode-Conditional Generation**: It generates images conditioned on specific modes, enhancing realism. - **Gaussian Mixture Models (GMMs)**: Fitted GMMs are used to model the multimodal distributions, improving the quality of generated images. **Tech stack:** MM-FM is built using PyTorch, TensorFlow, and Hugging Face for model training and inference. **Why it matters:** MM-FM is highly effective in generating high-quality multimodal images and videos, making it a valuable tool for researchers and practitioners in computer vision and image synthesis. **Repository:** The repository is hosted on GitHub at [GaoxiangLuo/MM-FM](https://github.com/GaoxiangLuo/MM-FM). **Primary language:** Python **GitHub stars:** 15 **Why it matters:** MM-FM is highly effective in generating high-quality multimodal images and videos, making it a valuable tool for researchers and practitioners in computer vision and image synthesis.

### Overview: RAGScore is a powerful tool designed to generate QA datasets and evaluate RAG systems. It simplifies the process of auditing and improving RAG models, making it accessible for developers and researchers alike. ### Key Features: - **Generate QA Pairs**: Quickly create QA datasets from your documents. - **Evaluate RAG Systems**: Assess the accuracy and correctness of your RAG system with just two commands. - **Privacy-First**: Ensures the privacy of your data through anonymization and encryption. - **Async Processing**: Supports both local and cloud-based evaluations. - **Multilingual Support**: Handles multiple languages seamlessly. - **Jupyter/Colab Integration**: Perfect for rapid development and iteration. - **Detailed Reports**: Provides comprehensive visual and tabular reports for easy analysis. ### Tech Stack: Python ### Why It Matters: - **Efficiency**: Streamlines the QA dataset generation and evaluation process. - **Accuracy**: Ensures high accuracy in QA pair generation and RAG system evaluation. - **Flexibility**: Supports both local and cloud-based evaluations, making it versatile for various environments. - **Accessibility**: Easy to install and use, suitable for both developers and researchers. ### Repository: - **Repository**: [HZYAI/RagScore](https://github.com/HZYAI/RagScore) - **GitHub Stars**: 15 - **GitHub Description**: The "1-Minute RAG Audit" — Generate QA datasets & evaluate RAG systems in Colab, Jupyter, or CLI. Privacy-first, async, visual reports. ### Installation

Overview: This repository contains the dataset, evaluation scripts, and baselines for the SemEval-2022 Task 5: Multimedia Automatic Misogyny Identification (MAMI). It serves as a platform for participants to submit their models and evaluate their performance. Key features: - Dataset (10000 + 1000 memes about misogyny detection) - Evaluation script and baselines - Baselines grounded on deep representations of text, image content, and their concatenation - Baselines for Subtask A and Subtask B - Instructions for running baselines and computing performance measures Tech stack: Python Why it matters: This repository provides a structured environment for researchers and developers to contribute to the automatic detection of misogyny in multimedia content, contributing to the advancement of natural language processing and image understanding technologies.

Overview: This repository is a mechanistic interpretability study focusing on the stability of induction heads in a 2-layer attention-only Transformer during domain adaptation from prose to structured Python code. It tracks the behavior of the induction scores and activation patching attribution across multiple seeds and checkpoints to understand how the model's internal circuits adapt or change. Key features: - Induction scores - Activation patching attribution - Multi-seed experiments with checkpoint analysis Tech stack: Python Why it matters: This study aims to understand whether induction circuits persist, adapt, or undergo phase changes when the model shifts domains, providing insights into the stability of the model's internal circuits during fine-tuning.

Overview: Fashion Sence AI is an AI-powered web application that enhances user experience by allowing image-based and text-based search for visually similar fashion products, personalized outfit suggestions, and trend inference. It utilizes CLIP, FAISS, and Gemma-3B for efficient image and text embeddings, and Streamlit for the frontend interface. Key features: - Image-based search for visually similar products - Text-based search for queries like "oversized hoodie, floral print dress" - Personalized outfit suggestions based on browsing history - Intelligent outfit completion using LLMs Tech stack: Streamlit, CLIP, FAISS, Gemma-3B, Hugging Face API Why it matters: This app serves as a virtual fashion assistant, personal stylist, or foundation for e-commerce AI integration, offering a more intuitive and personalized shopping experience. Repository: GitHub - MohitGupta0123/Fashion-Sense-AI GitHub stars: 15 GitHub description: AI-powered fashion visual search and personalized styling assistant using CLIP, FAISS, and Gemma-3B in Streamlit.

Overview: This repository contains the notebooks and projects code for a Generative AI for Backend Devs online bootcamp. It serves as a guide for students to learn and practice Generative AI concepts in the context of backend development. Key features: - Generative AI projects - Backend Devs training - Comprehensive guide Tech stack: Jupyter Notebook Why it matters: This guide is essential for students looking to enhance their skills in Generative AI and apply it effectively in backend development, making it a valuable resource for anyone interested in this field.

Overview: MCP-RAGAnything is a Python-based API and MCP wrapper that exposes a REST API and MCP server for document indexing and knowledge-base querying, powered by RAGAnything and LightRAG. It supports two retrieval pathways: a graph-based LightRAG pipeline and a classical RAG pipeline using multi-query generation with LLM-as-judge scoring. Files are retrieved from MinIO object storage and indexed into a PostgreSQL-backed knowledge graph. Key features: - Multi-modal RAG service - REST API and MCP server - Graph-based LightRAG pipeline - Classical RAG pipeline - Multi-query generation with LLM-as-judge scoring - Files retrieved from MinIO - Indexing into PostgreSQL-backed knowledge graph - Isolated projects via working_dir Tech stack: Python Why it matters: MCP-RAGAnything provides a robust and scalable solution for document indexing and knowledge-base querying, leveraging advanced retrieval techniques and powerful indexing capabilities. Its integration with RAGAnything and LightRAG makes it highly effective for applications requiring advanced search and retrieval functionalities. Repository: SoluDevTech/mcp-raganything GitHub stars: 15 GitHub description: API/MCP wrapper for RagAnything Branch: `feat/dual-auth-rls-llm-peruser` Breaking changes: Legacy `API_KEY` and `OPEN_ROUTER_API_KEY` for chat/embeddings are deprecated, replaced with JWT/OIDC via Logto or per-user API keys shared with composable-agents.

Overview: Ssebowa is a free and open-source Python library designed to provide generative AI models, including large language models (LLM), visual language models (VLLM), image generation and customized fine-tuning models, and video generation models. Key features: - ssebowa-llm: A large language model for text generation. - ssebowa-vllm: A visual language model for visual understanding. - ssebowa-imagen: An image generation and customized fine-tuning model. - Ssebowa-vigen: A video generation model. Tech stack: Python. Why it matters: Ssebowa offers developers a comprehensive suite of generative AI tools, enabling them to easily generate text, translate languages, create creative content, personalize images, and answer questions informatively. The library's flexibility and ease of use make it a valuable resource for anyone looking to integrate AI into their projects. GitHub stars: 15 Repository: SsebowaDisan/ssebowa

### Overview: VideoEval-Pro is a robust and realistic long video understanding benchmark designed for evaluating models on open-ended, free-form questions. This repository contains the evaluation code for the accepted paper at TMLR 2026, providing a comprehensive dataset and evaluation framework. ### Key Features: - Robust and realistic long video understanding benchmark. - Contains open-ended, short-answer QA problems. - Constructed from four existing long video understanding MCQ benchmarks: Video-MME, MLVU, LVBench, and LongVideoBench. - Data available on HuggingFace: [VideoEval-Pro](https://huggingface.co/datasets/TIGER-Lab/VideoEval-Pro). ### Tech Stack: - Python ### Why it Matters: This benchmark is crucial for evaluating the performance of models in understanding and answering open-ended questions from long videos, providing a realistic and robust evaluation environment for researchers and developers in the field of video understanding and language models. ### Repository: - Repository: TIGER-AI-Lab/VideoEval-Pro - GitHub Stars: 15 - GitHub Description: VideoEval-Pro: Robust and Realistic Long Video Understanding Evaluation [TMLR26] ### Primary Language: - Python ### GitHub Description: This repository contains the evaluation code for the [VideoEval-Pro](https://arxiv.org/abs/2505.14640), which has been accepted to TMLR 2026. ### Dataset Introduction: VideoEval-Pro is a benchmark dataset containing open-ended, short-answer QA

### Overview: Tahsine/kaggle-llm-server is a free LLM server running on a Kaggle GPU, accessible from any OpenAI-compatible client without needing a local GPU or paid plan. It enables users to run a multimodal LLM (text + images) on a free Kaggle GPU and connect to it from any machine. ### Key features: - **Multimodal LLM Server**: Runs a Qwen3.6-35B-A3B model with a 22.4 GB binary. - **Kaggle GPU Access**: Free GPU access from Kaggle, accessible via a Cloudflare tunnel. - **OpenAI-Compatible API**: Standard OpenAI-compatible API endpoint for interaction. - **Cloudflare Tunnel**: Public URL for easy access from any client. - **Model Compilation**: Compiles llama.cpp binaries for optimal GPU control and reduced memory overhead. ### Tech stack: - **llama.cpp**: Inference engine with CUDA support. - **Cloudflare Tunnel**: Public URL for easy access from any client. - **Kaggle**: Free GPU environment for running the server. ### Why it matters: - **Free Access**: No need for local GPU or paid plan, making it accessible to everyone. - **OpenAI Compatibility**: Works with any OpenAI-compatible client, ensuring seamless integration. - **Optimized Inference**: More precise GPU control and reduced memory overhead compared to wrappers. ### GitHub stars: 15 ### Repository: Tahsine/kaggle-llm-server

Overview: This advanced AI-powered conversational agent leverages the Llama 3.2 model and Phidata framework, offering enhanced reasoning, natural language interaction, and tool integration capabilities. Key features: - **Reasoning Capabilities**: Handles complex questions and provides logical responses. - **Tool Integration**: Access to DuckDuckGo for quick searches and a Calculator for basic arithmetic operations. - **Structured Output Responses**: Ideal for applications requiring organized data. - **Interactive Playground Interface**: Easy and user-friendly environment for interacting with the AI agent. - **Logging**: Comprehensive logging for debugging and monitoring. Tech stack: Python. Why it matters: This project is designed to provide a highly interactive and versatile AI solution, suitable for various conversational AI applications. Repository: TheSethRose/Agent-Chat GitHub stars: 15 GitHub description: An advanced AI-powered conversational agent leveraging the Llama 3.2 model and Phidata framework.

Overview: This repository contains a collection of miscellaneous codes and writings related to MLOps, focusing on various aspects such as AI as a service, AWS, Kubernetes, large language models, and ML serving. Key features: - AWS Lambda - SageMaker HuggingFace toolkits - Serverless inference with SageMaker - DVC for data management - Basic and troubleshooting examples for Kubernetes - Large language models (LLMs) including document-oriented agents, RAG, peft, fsdp_qlora, llama_index, TensorLLM, and prompt engineering - ML serving with BentoML, Triton Server, FastAPI, Flask, Ray, and Ray-serve Tech stack: Jupyter Notebook Why it matters: This repository provides practical examples and tutorials for various MLOps-related tasks, making it useful for developers looking to enhance their skills in deploying and managing machine learning models efficiently. Repository: YeonwooSung/MLOps Primary language: Jupyter Notebook GitHub stars: 15 GitHub description: Miscellaneous codes and writings for MLOps README excerpt: # MLOps ## Table of Contents - [AI as a Service](./ai-as-a-service/): simple AI-based service with serverless (AWS Lambda) - [AWS](./AWS/) - [Lambda](./AWS/Lambda/) * [invoke sagemaker with aws lambda](./AWS/Lambda/python/invoke_sagemaker.py) - [SageMaker](./AWS/SageMaker/) * [SageMaker H

### Overview: MEXA: Towards General Multimodal Reasoning with Dynamic Multi-Expert Aggregation MEXA is a model designed to tackle general multimodal reasoning tasks by dynamically aggregating the insights from multiple experts. This project aims to enhance the ability of models to understand and reason across different modalities, making them more versatile and effective in various applications. ### Key Features: - **Dynamic Multi-Expert Aggregation**: MEXA integrates insights from multiple experts to improve reasoning across multimodal data. - **Generalization**: The model is designed to generalize well across different tasks and datasets, making it applicable to a wide range of applications. - **Efficient Inference**: The inference script provided is straightforward and easy to use, making it accessible to developers and researchers alike. ### Tech Stack: Python MEXA is built using Python, leveraging the flexibility and scalability of this language to ensure efficient and robust performance. ### Why It Matters: MEXA is a significant advancement in multimodal reasoning, offering a more comprehensive and adaptable approach to understanding complex data from multiple sources. Its ability to dynamically aggregate insights from multiple experts makes it particularly useful in scenarios where traditional methods struggle to provide accurate and comprehensive reasoning. Repository: Yui010206/MEXA GitHub stars: 15 GitHub description: [EMNLP 2025 Findings] MEXA: Towards General Multimodal Reasoning with Dynamic Multi-Expert Aggregation

Overview: This project is a Python-based tool designed to import buildings into OpenStreetMap (OSM) using AI and orthophoto imagery. It utilizes a fine-tuned MobileNetV3Large model to classify the accuracy of BDOT10k information and verifies historical OSM data. Key features: - Importing buildings into OSM - Utilizing orthophoto imagery for building preprocessing - Fine-tuning a MobileNetV3Large model for building classification - Verifying historical OSM data for previously deleted buildings Tech stack: Python Why it matters: This tool simplifies the process of importing buildings into OSM, leveraging AI and high-quality imagery to ensure accuracy and completeness of the data, making it valuable for urban planners and data enthusiasts alike.

Overview: StableFused is a toy library designed to experiment with Stable Diffusion, inspired by various sources. It aims to help the developer learn about generative models in general. Key features: - Text-to-Image - Image-to-Image Tech stack: Python Why it matters: This library provides a fun and educational way for developers to experiment with generative models, inspired by popular projects like 🤗 diffusers and others. Repository: a-r-r-o-w/stablefused GitHub stars: 15 GitHub description: StableFused is a toy library to experiment with Stable Diffusion inspired by various sources.