Oct 2020 – Feb 2021 · 4 mos · Los Angeles, California, United States

• As a Machine Learning Intern at Easley-Dunn Productions I worked on machine learning techniques for video style transfer to create storytelling experiences for video games.

Jul 2020 – Sep 2020 · 2 mos · Pittsburgh, Pennsylvania, United States

• I worked on the Memory Management team at Qeexo. My responsibilities included making the Qeexo AutoML platform memory efficient in order to deploy it on embedded devices with limited computational and memory resources. My work involved conducting memory profiling for the AutoML code pipeline and proposing and implementing architectural code changes to remove redundant memory utilization. I also worked on adding functionality to the platform, including performance metrics and visualizations.

May 2019 – Aug 2019 · 3 mos · San Francisco Bay Area

• I was responsible for coding an end-to-end, distributed framework for automated Neural Network Compression for image segmentation tasks. The framework modelled compression as a policy optimization for a Markov Decision Process which searched in the architecture space of smaller networks, using a function of compression ratio and performance loss as reward. Exploration of each state by done an asynchronous advantage actor-critic agent and involved training of a smaller network with knowledge distillation using soft output matching. The framework achieved a 10.4x compression in the number of parameters for Chest Cavity Segmentation and 9.9x compression for Lung Segmentation tasks when compressing a U-Net architecture on GE Lung X-ray dataset with minimal loss in performance. I also generalized the framework to compress different Convolutional Neural Network architectures provided as input and made it end-to-end to be able to train the original network as well.

Jun 2017 – Aug 2017 · 2 mos · Stanford, California

• I worked on Deep Material Prediction for automated Robotic arms using tactile information. The aim was to automate the identification of materials by robotic arms via tapping on or scratching objects. I trained LSTM models with Attention to use accelerometer data from tapping experiments to achieve 72.3% accuracy. The data was processed using Wavelet decomposition and spectrograms.

May 2016 – Jul 2016 · 2 mos · Bangalore

• I worked on Scalable Downstream Impact Estimation to estimate the impact of initial customer purchases on future interaction with the Amazon marketplace. I created models for the same using Propensity Modelling, Attribution Theory and Linear Regression while dealing with big data using Apache Pig. I was also responsible for devising innovative filters to decide relevant products on which the downstream impact was modelled, and to combine similar products into one category for more efficiency.