Medtronic

1. Simulation and Training Workflow for 3D Soft Tissue Registration:
• Enhanced training codebase with optimized distributed training (utilizing in-house HPC) and asynchronous data loading techniques for improved efficiency and scalability of a 3D U-Net model while reducing training time by 85%.
• Simulated ground truth organ deformations using finite element analysis and Bio-Mechanical models for U-Net pre- training. Implemented a visualization script to interpolate output predictions in 3DSlicer.
• Reduced 3D registration error by 35%, by integrating multi-headed self and cross attention layers in vanilla 3D U-Net.
2. Long-Range 3D Pixel Tracking Workflow:
• Assessed Transformer and Optical Flow-based tracking algorithms for robust key point tracking on combination of open source and proprietary data, facilitating the registration of 3D models on soft tissue organs.
• Generated 3D point clouds from stereoscopic data using RAFT-Stereo for stereo reconstruction and depth estimation, incorporating camera intrinsic parameters.
• Accelerated long range 3D pixel tracking by creating a custom data loader, reducing memory usage by 75%.
• Collaborated within an Agile framework, contributing to bi-weekly sprints and weekly standup meetings to ensure seamless team coordination and project progress.

Mitacs

• Developed a pipeline for Driver Distraction Detection leveraging 3 SOTA Deep CNNs and designed a Dynamic Ensemble Model with a VGG Based Autoencoder improving combined performance of the baseline models.
• Integrated evaluation pipeline with Grid and Average Weighted ensembles using the weights and biases library.
• Addressed bias through data balancing, augmentation, and hyperparameter tuning to enhance generalization.
• Performed A/B Testing to select and assess the architectures for baseline models and its parameters.
• Improved model robustness with regularization and dropout, achieving benchmark classification accuracy of 89.13%.
• Published article in IEEE Journal: Driver Distraction Detection
• Tools and Frameworks: Python, NumPy, OpenCV, Keras, Tensorflow, Scikit-learn, Weights and Biases

Northeastern University

• Collaborated with professor to develop and refine course materials, and actively contribute to the planning and execution of lectures and programming assignments.
• Conducted weekly office hours to provide one-on-one assistance to students clarifying course concepts. Assessed student assignments and projects, providing constructive feedback.
• Assisted students with topics involving Genetic Algorithms, Hidden Markov Models and Graph based searching.
• Tools: Python, Pytorch, Numpy, Scikit-learn, Jupyter

Languages

Python

C++

SQL

Java

Cuda

Scala

Libraries

NumPy

Pandas

OpenCV

VTK

scikit-learn

matplotlib

Frameworks

Flask

Hadoop

Spark

Keras

TensorFlow

PyTorch

Other

Git

AWS

Jupyter

3D Slicer