•Worked on developing an application called Thornton Tomasetti Damage Detector (T2D2), which would be used to detect damages on buildings and structures through drones based on Deep Learning and Computer Vision.
•Trained a variety of deep learning models, including Faster R-CNN, Mask R-CNN and YOLO, and compared their results to choose the best model for detecting damages.
•Performed data cleaning, data labeling and data exploration to minimize training loss and optimize accuracy.
•Created an automated pipeline for damage detection using additional machine learning classifiers to improve prediction accuracy.
•Worked on optimizing the models for faster inference using techniques like pruning, constant folding and folding batch norms, resulting in a 35% increase in inference speeds.
•Helped create an online inference server to deploy the models using TensorFlow Serving with Docker, Google Cloud services and Kubernetes.
