Projects

Overview: ChessApp is an Android application designed for detecting and classifying chessboard images and chess pieces using deep learning techniques. Key features: - Chessboard detection - Chess piece classification - Virtual board visualization - Collaboration with Simone Mattioli Tech stack: Java Why it matters: This project showcases the application of deep learning and computer vision in detecting and classifying chessboard images and chess pieces, demonstrating the effectiveness of neural networks in real-world scenarios. Repository: MatteoLucchi1998/ChessApp Primary language: Java GitHub stars: 13 GitHub description: Deep Learning and Computer Vision application for chessboard detection and chess pieces classification. Datasets: Created datasets of 12.500 and 110.600 images Models: Generated up to 13 models using different approaches and training techniques All checkpoints: Available for download at <a href="https://drive.google.com/drive/folders/1Tp4o3D6_EMaBr4yM0hf4jzWkTPj1ZNeC">here</a>

Overview: EduMate RAG is a conversational question-answering service that retrieves answers from indexed course materials, powered by LangChain, Qdrant Cloud, and Groq LLM inference. It supports both integrated app flow and standalone service modes. Key features: - Answers only from indexed course materials - Remembers context across questions - Semantic retrieval using Qdrant Cloud or local ChromaDB - AI-powered generation with Groq's 120B model - Multilingual support (Arabic and English) - Source attribution in every answer - Runs on Groq's free tier - Incremental indexing with file change detection, embedding cache, and parallel PDF processing Tech stack: Python 3.11, FastAPI 0.109+, LangChain, Groq, Qdrant Cloud / ChromaDB, PyPDF 4.0.1+, and python-dotenv. Why it matters: This service enhances the EduMate platform by providing accurate and contextually relevant answers, improving user experience and educational effectiveness.

Overview: This repository contains the code and dataset needed to fine-tune the Qwen 2.5 AI model to serve as a smart customer service assistant, utilizing Python and Unsloth on Google Colab. Key features: - Fine-tuning of the Qwen 2.5 AI model - Utilizing Python and Unsloth - Training and testing on a dataset - Saving the trained model Tech stack: Python, Unsloth, Google Colab Why it matters: This project demonstrates how to adapt an AI model for customer service tasks, showcasing the potential of fine-tuning existing models for specific applications.

Overview: This project is a simulator designed to explore and optimize batching strategies and load patterns in Large Language Model (LLM) inference. It provides various configurations and metrics for easy comparison and analysis. Key features: - Multiple batching strategies (Static, In-Flight, Chunked Context, One Prefill Per Batch) - Different load generation patterns (Batch, Concurrent, Request Rate) - Rich metrics visualization - Configurable batch sizes and request parameters - Queue growth analysis for long-running simulations Tech stack: Python Why it matters: This simulator helps developers and researchers understand and optimize the performance of LLM inference by providing a comprehensive tool for exploring different strategies and patterns. Repository: Muhtasham/llm-inference-simulator Primary language: HTML GitHub stars: 13

### Overview: LoomRAG is a multimodal retrieval- augmented generation system that leverages OpenAI's CLIP and Whisper models for seamless text, image, and audio search. ### Key Features: - **Cross-Modal Retrieval**: Search text to retrieve both text and image results using deep learning. - **Image-Based Search**: Search the database by uploading an image to find similar content. - **Embedding-Based Search**: Uses OpenAI's CLIP, Whisper, and SentenceTransformer's Embedding Models for embedding the input data. - **CLIP Fine-Tuning**: Supports custom model training with configurable parameters including test dataset split size, learning rate, optimizer, and weight decay. - **Fine-Tuned Model Integration**: Seamlessly load and utilize fine-tuned CLIP models for enhanced search and retrieval. - **Upload Options**: Allows users to upload images, PDFs, and audio files (including real-time recording) for enhanced interaction and intelligent retrieval capabilities. ### Tech Stack: Python ### Why it Matters: LoomRAG streamlines the process of finding relevant information by combining text, images, and audio, making it a powerful tool for AI-powered search and retrieval.

### Overview: GALLO-3XL is a high-quality image generation model powered by NVIDIA A100, designed to replicate the style of DALL-E 3 with a focus on space themes. This project showcases the capabilities of AI in creating visually stunning and realistic images. ### Key Features: - **Space Theme**: The model is specifically designed to generate images inspired by space, making it ideal for projects requiring a futuristic or cosmic aesthetic. - **Gradio Integration**: The model is integrated with Gradio, a Python-based framework for building user interfaces, allowing for easy deployment and accessibility. - **Zero GPU Processing**: The model can process images without requiring a GPU, making it suitable for environments with limited computational resources. ### Tech Stack: Python ### Why it Matters: Gallo-3XL leverages AI to create stunning, high-quality images, making it a valuable tool for artists, designers, and anyone looking to enhance their visual storytelling. Its ability to replicate the style of DALL-E 3 with a focus on space themes makes it particularly appealing for projects requiring a futuristic or cosmic aesthetic. Repository: PRITHIVSAKTHIUR/GALLO-3XL Primary language: Python GitHub stars: 13 GitHub description: High Quality Image Generation Model - Powered with NVIDIA A100 README excerpt: ...

Overview: This repository contains the source code for a chatbot designed to assist customers with clothing advice, website navigation, and store information at a clothing store. Key features: - Provides clothing advice and recommendations - Guides users on how to navigate the store's website - Uses advanced language processing to understand and respond to user queries Tech stack: Jupyter Notebook Why it matters: This chatbot is a valuable tool for enhancing customer experience and improving the efficiency of a clothing store's operations by providing personalized and accurate information through a conversational interface. Repository: QuangNguyen711/AutClothingChatbot Primary language: Jupyter Notebook GitHub stars: 14 GitHub description: PTIT's Major Project: Website Programming - This repo contains a chatbot for a clothing store. The chatbot acts as an employee with specific knowledge about clothing consultation, website support, and store information. README excerpt:

Overview: This kit provides a reproducible setup for deploying DeepSeek-V4-Flash on a 2× NVIDIA DGX Spark cluster, featuring vLLM TP=2, NVFP4 KV, DSpark speculative decoding, and 1M context. The repository includes configs, scripts, and systemd units for a quick start. Key features: - DGX Spark setup - QSFP cable and control host - vLLM TP=2 configuration - NVFP4 KV integration - DSpark speculative decoding - 1M context support - systemd self-heal Tech stack: Python Why it matters: This kit provides a reliable and reproducible environment for deploying DeepSeek-V4-Flash, ensuring consistent results and ease of deployment.

Overview: SwarnDB is a unified data layer for AI that combines vector and graph databases into a single, efficient engine. It simplifies the integration of vector and graph databases by treating them as the same object, eliminating the need for synchronization and glue code. Key features: - A single engine for vector and graph databases - Production-grade Rust implementation - Inbuilt vector and graph functionalities - Opt-in graph functionality with explicit edges and provenance - Composable hybrid queries for efficient similarity and graph traversal Tech stack: Rust Why it matters: SwarnDB simplifies the integration of vector and graph databases, making it easier to build AI applications that require both vector and graph capabilities. The single engine approach reduces complexity and improves performance by eliminating the need for synchronization and glue code.

Overview: This project demonstrates the implementation of text summarization using Python, deep learning, machine learning, and transformer techniques. It integrates Hugging Face's Transformers library and Langchain for advanced summarization capabilities. Key features: - Text summarization using deep learning and machine learning models - Utilization of the Transformer architecture for improved summarization quality - Integration with Hugging Face's Transformers library for advanced summarization techniques - Compatibility with Langchain for seamless integration with other AI tools Tech stack: Python, deep learning, machine learning, transformer, Hugging Face Transformers, OpenAI, Langchain Why it matters: This project showcases the practical application of cutting-edge AI techniques in text summarization, making it valuable for developers looking to enhance their text processing capabilities.

### Overview: This Rust project runs the Qwen3.8 27B language model on a single CPU with at least 8GB of RAM, achieving up to 7.4 tokens per second. It is a pure Rust implementation that avoids BLAS, GPU, and ML frameworks, making it highly efficient and compatible with minimal hardware requirements. ### Key features: - **Pure Rust Implementation**: The entire model is written in Rust, without the use of BLAS, GPU, or ML frameworks. - **Quantized GEMM/GEMV Kernels**: Hand-written AVX2+FMA intrinsics for the quantized GEMM and GEMV kernels, optimized for performance. - **Scalar Stages Optimization**: The scalar stages around the quantized kernels are moved to the host feature set, enhancing compatibility and performance. - **Single 16.5 GB GGUF File**: The entire model is stored in a single 16.5 GB GGUF file, mapped directly from disk, ensuring minimal overhead. ### Tech stack: Pure Rust. ### Why it matters: This project is designed to be highly efficient and compatible with minimal hardware requirements, making it suitable for a wide range of environments, including single-core CPUs with modest RAM. ### Repository: StepfenShawn/qwen_inference_in_rust ### Primary language: Rust ### GitHub stars: 13 ### GitHub description: Runs the Qwen3.8 27B LLM on a single CPU with >=8GB of RAM, delivering up to ~7.4 tok/s

### Overview: FlashTopK is a fast, exact, and deterministic fused top-k selection kernel designed for learned sparse attention training and inference. It significantly improves performance compared to dense scoring methods, making it highly efficient and reliable. ### Key Features: - **Fast Performance:** FlashTopK is 1.7-3.3x faster than dense scoring + FlashInfer top-k on trained indexers and 1.2-2.0x faster on untrained indexers. - **Robustness:** It stays fast on both trained and untrained indexers, with adaptive key pruning to decide whether to skip keys based on a score bound. - **Exact Selection:** The true top-k selection, ensuring that rows sampled by the threshold miss are rescored on the GPU. - **Deterministic Output:** The same bits are produced every time, with ties at the k-th score preserved by keeping the smallest key indices. - **Context Parallelism Compatible:** Each row's threshold comes from a strided sample of its own window, allowing for efficient context parallelism. ### Tech Stack: Python, PyTorch, Triton, CuTe DSL (nvidia-cutlass-dsl) 4.5+ ### Why It Matters: FlashTopK is the fastest open-source implementation of fused top-k selection, making it highly efficient and reliable for training and inference tasks involving learned sparse attention.

### Overview: The Model-to-NPU pipeline for Qualcomm Snapdragon is a groundbreaking project that enables the execution of Stable Diffusion XL (SDXL) on-device, specifically on the Snapdragon 8 Elite (SM8750) and Hexagon V79 HTP NPU. This project aims to eliminate the need for cloud servers, Termux, rooting, or Python runtime, making it accessible to a wide range of users. ### Key features: 1. **Native NPU Execution**: The project runs the SDXL pipeline natively on the Qualcomm Hexagon NPU, achieving high performance and low latency. 2. **Unified UNet Model**: The UNet model is no longer split into separate encoder and decoder halves, eliminating bottlenecks and improving efficiency. 3. **Dynamic Resolution Engine**: The engine supports arbitrary aspect ratios and resolutions, enabling seamless scaling without recompiling models or context switching. 4. **No Root, No Termux, No Python**: The project is designed to be executed without rooting, Termux, or Python runtime, making it accessible on any Android device. 5. **Standalone Android APK**: A standalone Android application is provided, allowing users to run the pipeline directly on their device. ### Tech stack: The project utilizes the following technologies: - **Snapdragon 8 Elite (SM8750)**: The SoC on which the pipeline runs. - **Hexagon V79 HTP NPU**: The Qualcomm Hexagon NPU used for execution. - **QNN (Qualcomm Neural Network)**: The neural network system

Overview: This project is an AI tool designed to analyze the layout grid systems of posters, allowing users to detect and view the grid system, export it as SVG guides or CSS Grid/Tailwind code, and even export it as SVG guides or CSS Grid/Tailwind code. The tool is model-agnostic, using a multimodal vision model to analyze grid systems, elements, keylines, and color palettes. Key features: - Vision-model analysis of grid systems, elements, keylines, and color palettes. - Web UI with live grid overlay, layer toggles, and an export dialog. - Export options for SVG grid guides (Figma/Illustrator ready) or CSS Grid/Tailwind code. - CLI (`npm run analyze -- ./poster.jpg --summary`) for no server required. - OpenAI-compatible multimodal model support. - Robust pipeline with timeout, retry, schema validation, normalization, and in-memory cache. - Agent-ready, ships as an installable skill (`SKILL.md`) and `AGENTS.md`. Tech stack: HTML. Why it matters: This tool simplifies the process of analyzing and understanding the layout of posters, making it easier for designers and developers to work with grid systems and optimize their designs. Repository: Ygria/poster-grid-inspector GitHub stars: 13 GitHub description: Model-agnostic AI poster grid inspector: detect a poster's layout grid system with a vision model, view it overlaid, and export SVG guides / CSS Grid / Tailwind code. Ships as an installable agent skill.

Overview: This project fine-tunes the T5 model on the SQuAD dataset to generate questions. Key features: - Fine-tuning T5 model - SQuAD dataset integration - Question generation capability Tech stack: Python Why it matters: This project demonstrates the application of T5 fine-tuning on a practical dataset, showcasing its potential for generating questions from text data.

Overview: AV-Phys Bench is a comprehensive benchmark designed to evaluate how well audio-video generation models understand and preserve physical commonsense as scenes evolve over time. It includes three categories: steady state, event transition, and environment transition, with an Anti-AV-Physics subcategory to test generative physics knowledge. Key features: - Comprehensive dataset for evaluating physical commonsense in joint audio-video generation - Seven generators evaluated by human raters, an MLLM-as-judge baseline, and the AV-Phys Agent - Live leaderboard, video gallery, and per-prompt rubric details on the project page Tech stack: Python Why it matters: This benchmark is crucial for advancing the field of audio-video generation and understanding how models interpret physical reality, providing valuable insights into the limitations and capabilities of current AI systems. Repository: ZijunCui02/AV-Phys Primary language: Python GitHub stars: 13 GitHub description: Project page for AV-Phys Bench: Do Joint Audio-Video Generation Models Understand Physics?

Overview: This project is a computer vision system designed to detect and recognize license plates in images or video streams. Key features: - Detects license plates from video streams. - Extracts alphanumeric characters from the detected license plates. - Associates detected plates with their respective vehicles. - Utilizes a custom-trained license plate detection model. - Reads license plate contents using the PaddleOCR English pretrained model. - Tracks vehicles using ByteTrack algorithm. Tech stack: Python Why it matters: This project is significant for applications that require accurate and efficient detection of license plate information, such as traffic monitoring, license plate recognition systems, or automated license plate recognition for security purposes. Repository: ahmedshafiq12/License-Plate-Recognition-System Primary language: Python GitHub stars: 13 GitHub description: The License Plate Recognition (LPR) System is a computer vision project designed to detect and recognize license plates in images or video streams.

Overview: LLM Sentinel showcases NVIDIA's NeMo-Guardrails and LangChain for enhancing the safety of Large Language Models (LLMs). Key features: - LLM Sentinel integrates NVIDIA's NeMo-Guardrails and LangChain for improved LLM safety. - It utilizes the Llama3:8b model for development and the Llama3:70b model for production environments. - The project is compatible with Python 3.9 and above, running on macOS and Ubuntu Linux. Tech stack: Python Why it matters: This project provides a practical demonstration of how to enhance the safety of LLMs using cutting-edge AI tools, making it valuable for developers and researchers in the field.

### Overview: ComfyUI Module is a one-click portable installer for ComfyUI, a high-performance AI art generation tool. It eliminates the need for Docker, Python, Git, or admin rights, automatically downloading and isolating everything into a single, relocatable folder. ### Key Features: - **Multi-GPU Support**: Automatically detects and runs ComfyUI on multiple GPUs simultaneously, allowing for parallel processing. - **Multi-Instance Management**: Easily manage multiple ComfyUI servers with one-click controls, each with its own GPU, port, VRAM mode, and log prefix. - **101 Pre-Built Models**: Covers all default ComfyUI workflows, including Flux 1.x, SDXL, SD 3.5, SD 1.5, LTX-2, HunyuanVideo, Wan 2.1, and more, with automatic downloads and sorting by category and VRAM tier. - **16 Curated Custom Nodes**: Essential and recommended extensions, installed with automatic dependency handling. - **SageAttention**: One-click install for significantly faster attention computation, especially beneficial for video generation. - **Zero Dependencies**: No Python, Git, CUDA toolkit, Docker, or admin rights required. Everything is downloaded and configured into a portable folder. ### Why This Exists: - Eliminates the need for multiple installations, configurations, and dependencies. - Automatically downloads, configures, and isolates everything, ensuring no system changes or admin rights required. - Provides a fully relocatable, zero-dependency installation. ### Tech Stack: - Python

### Overview: This project optimizes Qwen3.5 fine-tunes on NVIDIA DGX Spark, achieving a 7.67× LoRA speedup and an 8.35× Full Fine-Tuning speedup compared to renting a H100 GPU. The loss curve is verified to be lossless within BF16 precision, and the training job runs on real multimodal data with a wall-clock parity to a rented H100 GPU in just 20 hours. ### Key features: - **LoRA Speedup:** 7.67× faster than the stock Unsloth. - **Full Fine-Tuning Speedup:** 8.35× faster than renting a H100 GPU. - **Lossless Training:** The loss curve is verified to be lossless within BF16 precision. - **Real Data Training:** The training job runs on real multimodal data with a wall-clock parity to a rented H100 GPU. - **Interactive Wizard:** The wizard handles model picker, data validator, training, and merge in three commands. ### Tech stack: Python ### Why it matters: This project demonstrates the potential of using your existing hardware to achieve significant speed improvements in fine-tuning models, making it cost-effective and practical for developers and researchers.