ML Engineer | Computer Vision (YOLOv8, U-Net, ResNet) | PyTorch • FastAPI • Docker | Ex-IBM AI Intern | Open to Work

projects

github

Divyanshu Sharma  • Peerlist

<h3>Overview</h3>Lumeo is a full-stack web application designed to enhance low-light images using advanced deep learning techniques. At its core, it employs a custom U-Net model built with PyTorch, specifically trained on the LOL dataset to effectively brighten dark photos. The application features a FastAPI backend for efficient inference and a React frontend that allows users to compare before and after images seamlessly. Additionally, Supabase is integrated for sharing results, ensuring a smooth and interactive user experience.<h3>Key Features</h3><ul><li>Deep Learning Enhancement: Utilizes a custom U-Net model to recover color and detail from near-pitch-black images.</li><li>Privacy-First: Offers local or server-based inference, ensuring data remains under user control.</li><li>Fast Inference: Delivers results in under 600ms on standard CPU hardware.</li><li>Modern Interface: Provides a clean, responsive React frontend for easy drag-and-drop enhancement.</li></ul><h3>Technical Details</h3>The application is built using a robust tech stack that includes React for the frontend, FastAPI for the backend, and PyTorch for machine learning tasks. Deployment is handled via Vercel and Hugging Face Spaces, with Docker ensuring consistent environments. The use of Tailwind CSS and Framer Motion enhances the UI/UX, while Axios facilitates HTTP requests. The model's performance is optimized with a combined loss function and advanced optimization techniques.

Lumeo

BananaClock is an innovative AI-powered web application designed to minimize food waste by analyzing the ripeness of bananas. Utilizing advanced computer vision models, YOLOv8 and ResNet50, BananaClock accurately detects bananas in images and classifies their ripeness into five distinct stages: Green, Yellow-Green, Yellow, Yellow-Brown Spots, and Brown. This classification helps users predict the shelf life of their bananas, providing a countdown of days until they spoil.Key Features:<ul><li>Dual-Model AI Pipeline: YOLOv8 detects bananas, while ResNet50 classifies their ripeness.</li><li>Shelf-Life Prediction: Offers an accurate "Days Until Bad" countdown based on ripeness stage.</li><li>Smart Storage Tips: Provides personalized advice to extend freshness, such as refrigeration or separation.</li><li>Interactive Highlighting: Allows users to click on ripeness stats to highlight corresponding bananas in their photos.</li><li>Recipe Suggestions: Offers instant recipes for overripe bananas, like Banana Bread and Smoothies.</li><li>PWA Support: Can be installed as a mobile app with offline capabilities.</li><li>Social Sharing: Generates beautiful stat cards for sharing on Instagram and Twitter.</li></ul>BananaClock is open source and available under the MIT license, encouraging community contributions to further enhance its capabilities.

BananaClock

A Coffee Machine Simulator with a Python GUI, utilizing MySQL for data storage.

Coffee Machine Simulator

This project is an intelligent image classification system that uses deep learning to distinguish between cats and dogs in photographs. Built with TensorFlow and Keras, it demonstrates the practical application of Convolutional Neural Networks (CNNs) for computer vision tasks.

The CNN model features a sophisticated architecture designed for binary image classification, processing images at 256x256 pixel resolution with RGB channels. It utilizes convolutional layers with max pooling for feature extraction, followed by dense layers for classification decisions. The model employs sigmoid activation for binary output, optimized with Adam optimizer and binary crossentropy loss function.

The model achieves impressive performance metrics with approximately 85% validation accuracy, demonstrating balanced performance across both classes with 85% precision, 84% recall, and 84% F1 score. The training utilized the popular Dogs vs Cats dataset from Kaggle, comprising approximately 25,000 images, providing substantial data for robust learning without class bias.

The project features a user-friendly web application built with Streamlit and deployed on Hugging Face Spaces, making it globally accessible without technical setup requirements. The interface supports multiple input methods including drag-and-drop file uploads, image pasting, and URL-based linking. Users receive immediate feedback with clear predictions and confidence scores displayed as intuitive progress bars.

The development follows best practices with separate Jupyter notebook for model training and clean deployment scripts. The modern Python stack includes NumPy for computations, Pillow for image processing, and Requests for web inputs. This project showcases the complete machine learning pipeline from data preprocessing and model training to web application development and cloud deployment, making it an excellent demonstration of end-to-end ML application development.

Image Classification of Cats vs Dogs using CN

Diwali sales data analysis with Python libraries like numpy, pandas, matplotlib, seaborn.

Diwali Sales Python Analysis

HR Analytics Dashboard to measure employee attrition due to different factors.

HR Analytics Dashboard

This project focuses on analysing Uber trip data from New York City to build a predictive model for trip demand. The dataset contains four key columns: dispatching base number, date, active vehicles, and trips. The goal of the project is to identify patterns, understand peak times for trips, and predict future Uber trip demand using machine learning techniques. The process is divided into several stages: Data Cleaning, Feature Engineering, Model Training, Evaluation, and Visualization.

Uber Trip Analysis

The IBM HR Analytics Employee Attrition & Performance project aims to understand employee attrition and performance using machine learning techniques. This project involves three main phases: data cleaning, data analysis, and model building.