Achieved 98% pixel-based f1-score to eliminate server costs by creating a real-time model to implement document detection at 10 FPS on mobile phones; used novel data augmentation techniques to create synthetic data which immproved the model score buy more than 20%.

• Developed a model on 3D point clouds of buildings created by using drone images to calculate total surface area using Segmentation; decreasing the cost and time of on-site surveys.• Improved f1-score with a deep learning model to 99% from 75% for Embedded System using time-series sensor data to detect a specific signal anomaly.• Grew roof segmentation accuracy by 5%; incorporated it to the code-base. Decreased the cost of on-site surveys by developing an end to end Computer Vision technique to detect changes in building structure over time using satellite images, cutting the cost of an in-person visit.• Directly supervised and trained a team of two machine learning engineers on various projects.

• Achieved 99% F1-Score by creating a DL pipeline to detect LCD characters and classify 4 Hexa- decimals; applied augmentation techniques after designing the dataset from scratch; secured funding for University of Trieste’s Talent Fellowship Programme.• Developed an apparel recommender system for Trendage. Designed an API with a classification and recommendation response time of 50 ms; classified apparel based on labels and categories using data from a crowd-funded application.• Designed an algorithm with PSPNet to decrease plant damage by enabling a drone-mounted camera to segment and identify affected areas in plants for early disease detection. • Real-Time Traffic Light Detection using Deep Learning for self-driving cars using retina-net achieving a detection f1-score of 85%.• Developed an Algorithm for Utility Poles detection using segmentation(PSPNet) followed by the largest contour filtering achieving an f1-score of 93%.

• Achieved 98% f1-score at 30 FPS by developing a real-time detection(nadir view) and recognition model (slanted view) to classify 44 different vehicle types. The data was collected and labelled for local vehicles.• Collaborated with a PhD student to classify and segment retinal vessels and eye abnormalities using Fully Convolutional Densenet; improved segmentation accuracy by 5% compared to U-Net.