I am an incoming PhD student at the University of Massachusetts Amherst. I will be working with Dr. Scott Neikum and the SCALAR Lab. Prior to this, I did an MS in Computer Science at the University of Colorado Boulder. I was fortunate to be advised by Dr. Bradley Hayes and be a part of the Collaborative Artificial Intelligence and Robotics Labotatory (CAIRO Lab). My research is at the intersection of Human preference learning and AI safety. I am particularly intested in designing AI systems that are safe and better human aligned.

I completed my undergrad from Delhi Technological University in 2019, where I majored in 'Information Technology'. I have also worked as a Software Developer with Citicorp Services India Pvt. Ltd. before joining the master's program.

In my free time, I like to go for bike rides and runs. I also enjoy playing the guitar and exploring new music.

The work operationalizes an evaluation approach to determine the acceptability of human preference and obtain notions of bounded risk in learning from preferences. The threshold for risk is defined based on the distribution of expected rewards returned for the optimal policy and different user policies are evaluated for acceptability. Our results show that the approach can identify which policies are acceptable given our tolerance for risk. It is also able to remove failed demonstrations based on the evaluation.
Github   •   Thesis Paper

Task: A human and UR5e robot arm clearing cans from a table top

The robot maintains a belief over the human's goal, continually updating based on human’s motion at each time step. A POMDP framework is used to figure out the best action for the anipulator arm by reasoning over uncertainty in the human goal.
Github  •   Report

Developed an interactive grid world game to collect human demonstration data and implement inverse reinforcement learning to learn the reward function from the observed behavior of the human player.
Github   •   Report

A GAN-based latent fingerprint enhancement algorithm to improve the quality of fingerprint images while preserving the ridge structure (Used IIITD-MOLF Dataset)

NFIQ scores of enhanced images were three times lower than state-of-the-art approaches (around 1.88%)