I have worked with interesting people. Here are a few places where I have worked.

Microsoft Research, India

Mentors: Nagarajan Natarajan and Aditya Kanade
Project: NoFunEval: Funny How Code LMs Falter on Requirements Beyond Functional Correctness

• Developed benchmarks and evaluation methodologies for code generation with Large Language Models (LLMs), focusing on evaluation of non-functional requirements of the generated code.
• Extensively evaluated twenty-two code LMs to discover they falter on our benchmark and don’t sufficiently comprehend code they can otherwise edit.
• Building a capable coding LM to better comprehend and edit code.

Mentors: Nishanth Chandran, Dimitrios Dimitriadis, and Divya Gupta
Project: Fed-Encoder: A One-Shot Federated Learning Solution

• Led efforts to enhance the accuracy of global model trained in a challenging one-shot federated learning setup with clients trained on non-IID data.
• Achieved over 5% increase in accuracy compared to baselines in our most non-IID setting by combining client logits and encoder output from a pre-trained autoencoder.

Robotics Institute, Carnegie Mellon University

Mentors: Prof. Katia Sycara and Dr. Dana Hughes
Lab: Advanced Agent Robotics Technology Lab
Project: Explanation for Multi-Agent Search and Rescue Task

• Working on modeling the beliefs possessed by agents using a Multiple-Observer model thus interpreting the decisions via queries taken by them in a multi-agent team setting in a search and rescue task simulated in the Minecraft environment.
• Currently working on developing natural language descriptions of the observations from the metadata and developing other possible queries and memory representations.

Microsoft Research, New York City

Mentors: Olga Vrousgou and Pavithra Srinath
Team: Real World Reinforcement Learning Team
Project: Empirical Analysis of Privacy Preserving Learning

• Analyzed the effect of Aggregated Learning (i.e. saving only those features after training that have crossed a certain threshold of users) on the model learning capabilities of the Open Source library VowpalWabbit.
• Achieved competitive performance between the public model (does not have access to the user feature mapping) and private model (has access to the user feature mapping) in our empirical analysis of the privacy-preserving feature on the OpenML Click Prediction Dataset.
• Implemented two different approaches for the feature and compared the existing benchmarks performances of each, along with developing new unit tests and benchmarks for it.

Indian Institute of Science, Bangalore

Mentor: Prof. Shalabh Bhatnagar
Project: Reinforcement Learning for Scheduling Railway Lines

• Worked on evaluating widely used single-agent approaches, such as Dueling Deep Q Networks (DDQN) and Proximal Policy Optimization (PPO), using distributed scalable implementations from RLlib, to identify their shortcomings on the multi-agent Flatland Environment for efficient train scheduling.
• Reviewed the existing literature on Markov Decision Process (MDP) for Vehicle Rescheduling Problem and the algorithms developed to solve the problem specifically catering to the scheduling of trains.

Indian Institute of Technology Kharagpur

Mentor: Prof. Niloy Ganguly
Lab: Complex Networks Research Group
Project: Class Activation Maps for predicting Textual Entailment

• Experimented with Class Activation Maps in NLP to understand the words affcting the textual entailment prediction being made by the CNN model on the SNLI dataset consisting of 570k sentence pairs. Achieved a 60% test accuracy.
• Built a preprocessing pipeline for tweets on ebola for classification of their GloVe embeddings with Bidirectional LSTM.