Projects

### Overview: Madlab is an advanced AI development studio designed to streamline the process of dataset management, model training, and evaluation. It combines a robust Node.js/Express backend with a powerful Python-based training engine and a modern React frontend. ### Key Features: - **Training**: Supports LoRA/QLoRA fine-tuning with configurable rank, alpha, dropout; NEFTune for noisy embeddings; DoRA for better quality at the same rank; Config presets for quick tests, small, large models, and quality fine-tune; Job queue for multiple training runs; Checkpoint resume and graceful cancellation; Real-time loss chart and progress timeline. - **Dataset Management**: Multi-format import, dataset profiling, versioning, preview, validation rules, and LLM-powered augmentation. - **Evaluation**: BLEU/ROUGE metrics, LLM Judge, auto test case generator, and failure analysis. - **LLM-Powered Features**: Magic Import, dataset quality analyzer, hyperparameter advisor, and synthetic data generation. - **Infrastructure**: Health dashboard, audit logging, structured logging, project backup/restore, model lineage tracking, response caching, LM Studio health probe, and response caching. - **UI/UX**: Command palette, keyboard shortcuts, dark/light theme, toast notifications, and export bundle. - **Desktop App (Experimental)**: Electron packaging for standalone desktop app. ### Tech Stack: Node.js/Express backend, Python training scripts (PyTorch/Transformers), React frontend. ### Why It Matters: Madlab simplifies the complex process of AI model development, making it accessible

Overview: This project is a multi-user chatbot application built using Langchain and Pinecone in Next.js, providing comprehensive documentation on LangChain and vector databases like Pinecone. Key features: - Multi-user chatbot functionality - Integration with Langchain and Pinecone - Crawling of multiple sites - Unique username feature for conversation history management - MongoDB integration for database storage - Ability to specify different domains - Crawling of a single site (Lightning Tools) - Installation instructions for cloning the repository, setting up environment variables, and running locally Tech stack: TypeScript, Next.js, Langchain, Pinecone, Prisma, MongoDB, Able, LangChain Document Helper Why it matters: This project offers a comprehensive solution for developers looking to integrate LangChain and Pinecone into their applications, providing a robust multi-user chatbot system with advanced features like conversation history management and unique username functionality.

### Overview: Aster is a project aimed at providing blind and low-vision students with the ability to follow lecture videos without relying on screen readers, captions, or AI narrations. The goal is to bridge the gap where students like Meera, who are blind, are unable to fully participate in educational content due to the lack of visual information. ### Key Features: - **Blind-friendly video descriptions**: Aster automatically generates descriptions for videos, ensuring that blind and low-vision students can understand the content. - **Interactive learning**: Students can pause, rewind, and replay the video, enhancing their understanding and retention. - **Accessibility features**: Aster supports screen readers, captions, and transcripts, ensuring that all students can access the information. ### Tech Stack: - **Frontend**: React.js - **Backend**: Node.js - **Database**: MongoDB ### Why It Matters: The gap between students who can access educational content and those who cannot is significant, especially in regions where free education is the norm. By providing a solution that is both accessible and educational, Aster aims to democratize knowledge and empower individuals who might otherwise be left behind.

### Overview: GLM-5.3-Flash BIZ is a business-use intent framework designed for DGX Spark-class GB10 systems, focusing on three serving lines: 1.x, 2.x, and 3.x. This repository provides a comprehensive set of tools and documentation for developers to deploy GLM-5.3-Flash on these systems using TensorFold's native engine. ### Key features: - **TensorFold Engine**: Supports TP=2 and TP=3 with FP8 KV, text, tool calls, and structured output. - **Checkpoint Options**: Available for NVIDIA's pinned NVFP4 checkpoint (1.x), TensorFold checkpoint (2.x), and XSFP4 checkpoint (3.x). - **Setup Scripts**: Detailed setup instructions for each line, including text, images, video, and structured output. - **Documentation**: Comprehensive documentation and changelogs for each line, including version and changelog information. ### Tech stack: Python-based framework with support for Apache-2.0 code and weights fetched separately. ### Why it matters: GLM-5.3-Flash BIZ is designed for developers looking to deploy GLM-5.3-Flash on DGX Spark-class systems, providing a business-use intent framework with commercially usable licensing, pinned assets, and recorded checks for reversibility.

Overview: This repository contains the DeepSeek-V4.1-Flash-1017 model optimized for a single 256 GB M5 Ultra Mac Studio, utilizing TensorFold's native Metal engine. The model supports DSpark drafts, prompt caching, images, tools, and structured output, with a 1M context window and 70.7-104.9 token/s decode speed. Key features: - DSpark drafts - Prompt cache - Images - Tools and structured output Tech stack: Python Why it matters: This optimized model is designed for high performance on a single GPU, making it ideal for applications requiring fast text generation and prompt-based tasks. The use of TensorFold's native Metal engine ensures optimal performance and efficiency, making it suitable for both development and production environments.

### Overview: Omega-Forge is an unofficial training implementation designed to adapt VGGT-Omega (VGGT-Ω) to your own dataset. It provides a practical way to fine-tune VGGT-Ω without waiting for official training code to be released. Additionally, it includes a standalone 3D reconstruction visualizer that runs on any machine without the need for a GPU. ### Key Features: - **Training Implementation (`training/`)**: Adapts VGGT-Omega to your dataset using a JSONL manifest of posed RGB-D frames. Includes a held-out validation loop for control. - **Reconstruction Visualizer (`visualization/`)**: A standalone 3D reconstruction viewer that imports neither PyTorch nor VGGT-Ω, rendering any reconstruction on any machine without a GPU. ### Tech Stack: Python ### Why It Matters: - **Practical Adaptation**: Provides a way to fine-tune VGGT-Ω without waiting for official training code. - **Standalone Visualizer**: Offers a standalone 3D reconstruction viewer that can render any reconstruction on any machine. - **Improved Accuracy**: Finetuning results show a significant improvement in depth and pose accuracy across six benchmarks, including a 60% reduction in Absolute Relative Error (AbsRel) on held-out test. - **Image Resizing Issue**: Not stretching the frame in your data loader may actually cost more accuracy than fine-tuning will ever give back. This issue was an unexpected finding that turned out to be more important than any of the fine-tuning results.

### Overview: OMGM: Orchestrate Multiple Granularities and Modalities for Efficient Multimodal Retrieval is an implementation of a paper presented at the ACL 2025 Main Conference. The project focuses on multimodal retrieval, specifically addressing the challenge of efficient retrieval of multiple granularities and modalities. The implementation is built using Python and utilizes the Flash-Attn technology stack. ### Key Features: - **Multimodal Retrieval**: The project aims to efficiently retrieve information from multimodal data, such as text and images. - **Efficient Retrieval**: Utilizes advanced techniques to speed up the retrieval process, making it suitable for large-scale applications. - **Orchestration**: The system is designed to handle multiple granularities and modalities, ensuring flexibility and scalability. ### Tech Stack: - **Python**: The primary programming language used for the implementation. - **Flash-Attn**: A technology stack that enhances the performance of the retrieval process. ### Why It Matters: The project addresses a significant challenge in the field of multimodal retrieval, providing a solution that can be applied to various applications requiring efficient and accurate retrieval of information from multimodal data. The use of Flash-Attn technology ensures that the project is both efficient and scalable, making it valuable for researchers and developers looking to improve the performance of their multimodal retrieval systems.

Overview: This project explores the identity inference attack (IDIA) on the CLIP model, revealing whether individuals used for training can be identified through their images. The approach demonstrates high accuracy in identifying individuals trained on the Conceptual Captions 3M and LAION-400M datasets. Key features: - Identity Inference Attack (IDIA) - Large-scale experiments on CLIP models - Demonstration of high accuracy in identifying individuals used for training Tech stack: Jupyter Notebook Why it matters: The results highlight the need for stronger privacy protection in large-scale models and suggest that IDIAs can be used to prove unauthorized use of data for training and enforce privacy laws.

Overview: Dolphin is an open-source persistent memory infrastructure designed specifically for applications using large language models (LLMs) and AI agents. It combines vector search with a knowledge graph to enhance the learning and decision-making capabilities of these applications. Key features: - A Python SDK for integrating Dolphin into LLM applications. - Local storage capabilities using SQLite or Supabase for both local and server-side applications. - A shared memory feature for agents working on the same project, where each session's knowledge is available to subsequent sessions. - Support for local LLMs like Claude Code. Tech stack: Python, SQLite, Supabase, MCP, Ollama. Why it matters: Dolphin enhances the performance and efficiency of LLM applications by providing a persistent memory layer that allows for quick access to relevant information and knowledge graph-based reasoning. This makes it particularly useful in applications where the context and history of past interactions are crucial for decision-making and learning. Repository: DewashishCodes/dolphin GitHub stars: 14 GitHub description: Open-source persistent memory infrastructure for LLM applications.🐬 GitHub description excerpt: Open-source persistent memory infrastructure for LLM applications.🐬

### Overview: AIyu is a versatile shell-based framework designed to streamline the integration and execution of advanced AI technologies. It simplifies the process of building complex AI pipelines by providing a seamless interface for various AI components. ### Key Features: - **Prompt-to-Text (p2t)**: Translates text prompts into AI-generated responses. - **Text-to-Subtitle (t2s)**: Converts text into subtitles for video content. - **Subtitle-to-Text (s2t)**: Translates subtitles into text. - **Subtitle-to-Subtitle (s2s)**: Converts subtitles into subtitles. - **Text-to-Audio (t2a)**: Translates text into audio. - **Audio-to-Subtitle (a2s)**: Converts audio into subtitles. - **Audio-to-Text (a2t)**: Translates audio into text. - **Text-to-Code (t2c)**: Converts text into code. - **Code-to-Code (c2c)**: Compiles code. - **Audio-to-Audio (a2a)**: Enhances audio quality. - **Audio-to-Subtitle (a2s)**: Converts audio into subtitles. - **Text-to-Image (t2i)**: Generates images from text prompts. ### Tech Stack: Go ### Why It Matters: AIyu simplifies the integration of AI technologies into pipelines, making it easier for developers to leverage cutting-edge AI tools without the need for extensive coding. The framework supports multiple languages and is compatible with both Linux and macOS, making it accessible

### Overview: NemoANN is an SSD-based graph index for filtered approximate nearest neighbor search. It efficiently filters queries in memory before reading candidates from the SSD, significantly reducing I/O operations while maintaining high recall and throughput. The key feature is graph tunneling, where only relevant nodes are read from the SSD, optimizing performance even at low selectivity levels. ### Key Features: - **Graph Tunneling**: Reads only the relevant nodes from the SSD, significantly reducing I/O operations. - **Efficient Query Filtering**: Checks queries in memory before reading candidates, improving performance. - **High Recall and Throughput**: Maintains high recall and throughput even at low selectivity levels. - **Original Index Preservation**: Keeps the original index structure, making it 14x smaller on SSD compared to densified PipeANN-Filter. ### Tech Stack: - **C++**: The primary programming language used for the implementation. - **Linux**: Requires Linux with `io_uring` support. - **CMake**: Used for building the project. - **liburing**: A library for asynchronous I/O. - **tcmalloc**: A memory allocator. - **Python 3**: Required for experiment scripts. ### Why It Matters: NemoANN offers a significant performance improvement over existing methods, particularly in terms of I/O efficiency and throughput. It achieves 10x fewer SSD reads and 8.0x higher throughput compared to PipeANN with post-filtering, and even better performance gains at lower selectivity levels. This makes it highly relevant for applications requiring high recall and low I/O overhead

Overview: Nano Banana is an AI image generation plugin for Claude Code, powered by Google Gemini API. It allows users to generate and edit images directly within Claude Code using simple English prompts. Key features: - Text-to-image - Image editing - Style transfer - Multi-image composition - High-resolution output - Search-grounded generation - Support for various aspect ratios - Iterative editing Tech stack: Python Why it matters: Nano Banana simplifies the process of generating high-quality AI-generated images, making it easier for users to create professional-grade content without the need for complex setup or manual intervention. Repository: Ibrahim-3d/nano-banana-claude-plugin Primary language: Python GitHub stars: 14 GitHub description: Nano Banana - AI image generation plugin for Claude Code. 13 modes: text-to-image, editing, style transfer, inpainting, 4K, search grounding. Powered by Google Gemini API.

Overview: This project aims to develop a deep learning model for detecting pneumonia from x-ray images using a fine-tuned VGG-16 architecture. The goal is to reduce false negatives and improve overall accuracy, particularly in medical settings where false negatives can be fatal. Key features: - Fine-tuned VGG-16 model for pneumonia detection - Balanced dataset from Kaggle - Image augmentation using imgaug library - Training for 5 epochs - Evaluation metrics: precision and recall - Class activation maps for model interpretability Tech stack: Python (Jupyter Notebook), TensorFlow/Keras, Flask Why it matters: This project demonstrates the potential of deep learning in medical imaging, specifically in detecting pneumonia, which can significantly impact patient outcomes. The use of transfer learning and advanced preprocessing techniques ensures the model's effectiveness and reliability in real-world applications.

### Overview: ZeroGen is an innovative zero-shot multimodal controllable text generation model that leverages multiple oracles to generate text captions and control the generated content. This project is a PyTorch implementation of the research paper "ZeroGen: Zero-shot Multimodal Controllable Text Generation with Multiple Oracles" presented at NLPCC 2023. ### Key Features: 1. **Zero-Shot Generation**: ZeroGen can generate captions for images without any labeled data, making it highly versatile. 2. **Multiple Oracles**: Utilizes multiple oracles to enhance the quality and diversity of generated captions. 3. **Controllable Text**: Allows users to control the generated text by adjusting parameters like length, alpha, and beta, among others. ### Tech Stack: - **Python**: The primary programming language used for implementing the model. - **Transformers**: Used for text generation and control. - **NLTK**: For natural language processing tasks. - **Scikit-learn**: For data processing and model evaluation. - **Torch**: For deep learning frameworks and model training. ### Why It Matters: ZeroGen offers a significant advancement in the field of multimodal text generation by providing a zero-shot approach to captioning and text control, which is particularly useful in scenarios where labeled data is scarce or expensive to obtain. The model's ability to leverage multiple oracles makes it highly adaptable and effective across various domains. ### Repository: - **Repository**: ImKeTT/ZeroGen - **GitHub Stars**: 14 - **GitHub Description**: Official PyTorch

Overview: A generic frontend-only page for accessing various LLM APIs, designed by Jannchie. Key features: - Support for ChatGPT API. - Support for Open Router API. - Compatibility with all OpenAI-compatible platform APIs (Custom mode). - Support for chat history. - Displays reasoning process (Thinking). - Implemented text fade-in effect similar to Gemini. - Replica of Gemini-style chat UI. - Renders mathematical formulas, optimized for outputs from models like ChatGPT. - Supports code block copying and Markdown-formatted text copying. - Image Upload. Tech stack: Vue. Why it matters: This chat UI is designed to be a versatile frontend component for interacting with various language models, making it useful for developers and users alike. Repository: Jannchie/chat-ui Primary language: Vue GitHub stars: 14

### Overview: VecPuff is a vector database built on top of Amazon S3, inspired by the turbopuffer architecture. It is designed for learning and experimentation with scalable vector search architectures. ### Key features: - **Scalable Vector Search**: Built on S3 for efficient storage and retrieval. - **Log-Structured Architecture**: Uses WAL, compaction, and ANN index for fast search. - **Local Cache**: Caches frequently accessed files for improved performance. - **HTTP Server**: Runs on `http://localhost:3000` by default. ### Tech stack: Rust ### Why it matters: VecPuff simplifies the process of experimenting with scalable vector search architectures, making it accessible for learning and development.

Overview: This project fine-tunes a large language model (LLM) for EEG data classification, utilizing four movements (left hand, right hand, tongue, foot) to predict them from EEG signals. The model achieves better performance through fine-tuning and demonstrates the potential of LLMs in supervised learning. Key features: - Fine-tuned LLM for EEG data classification - Utilizes LLM for classification operations - Trained on gpt-4o model for improved performance - Performance metrics: Accuracy, F1 score, ROC-AUC Tech stack: Python, openai, mne Why it matters: This project showcases the potential of LLMs as supervised learning models, highlighting their capabilities in EEG data classification.

Overview: This project provides a minimal C++ framework for understanding and learning about the LLM inference pipeline. It includes key steps like config loading, weight loading, tokenizer, prefill/decode, sampling, and decoding, aiming for readability and modifiability. Key features: - Continuous batching service mode - Regression/invariant tests - Strong architecture/key/shape validation - CPU-first by default Tech stack: C++ Why it matters: It simplifies the learning process by providing a clear and editable pipeline, making it accessible for coursework, research, and self-study without focusing on peak performance. Repository: MarsPain/easy_llm.cpp GitHub stars: 14 GitHub description: A minimal C++ framework for learning and understanding the LLM inference pipeline.

Overview: This single-header C++17 library provides a set of 26 self-contained headers for integrating LLM (Large Language Model) features into C++ applications. These headers include streaming, retries, caching, cost estimates, RAG (Read-After-Write), structured JSON output, tool-calling agents, and more. The library is designed to be drop-in compatible with OpenAI and Anthropic APIs, making it suitable for developers shipping C++ applications without the need for a Python SDK or package manager. Key features: - **Streaming**: Support for real-time responses. - **Retries**: Automatic retry mechanisms. - **Caching**: Efficient caching of responses. - **Cost Estimation**: Real-time cost estimates. - **RAG**: Retrieval-Augmented Generation. - **Structured JSON Output**: Output in a structured format. - **Tool-Calling Agents**: Agents that can call external tools. - **Scrubbing PII and Keys**: Protection against exposing personal information. - **Cheaper Model Routing**: Optimized model routing for lower cost. - **Logging and Trace**: Detailed logging and tracing capabilities. - **Batch Prompts**: Support for batch processing. - **Mocking**: Mocking APIs for testing purposes. - **Images and Audio**: Support for image and audio processing. Tech stack: The library is built using C++17 and relies on the C++ standard library for most functionality. It also requires libcurl for API calls to OpenAI and Anthropic. Why it matters: This library simplifies integrating LLM features into C++ applications,

Overview: A modular, multilingual, and multimodal Retrieval-Augmented Generation (RAG) system tailored for the financial analysis of Public Investment Fund (PIF) annual reports. Key features: - Modular FastAPI backend - Vector database integration (Qdrant) - PDF ingestion and caching - Web frontend for interactive analysis - Prometheus metrics - Dockerized deployment Tech stack: Python Why it matters: This framework extends the design principles of M3DocRAG, specifically adapted for financial document understanding in both Arabic and English, making it a valuable tool for financial analysts and researchers. Repository: MohamedMostafa259/pif-multimodal-rag GitHub stars: 14