Advisor: Prof. Graham Neubig1. Building AI coding agents that can process multimodal data, including images, web-pages and PDFs.2. Aim to improve performance on benchmarks that evaluate web navigation, document understanding, and visual reasoning.

Title: Learning Task Agnostic Triggers for Auditing API-based LLMsBachelor's Thesis Project (nominated for Best Bachelor's Thesis Project Award)1. Developed an algorithm to audit the adversarial robustness of API-based LLMs (like GPT-3.5 Turbo).2. Deployed gradient-based adversarial attacks to learn input-agnostic and task-agnostic triggers that could significantly degrade LLM performance across various tasks like translation, sentiment analysis, question-answering, and natural language inference.Advisors: Prof. Somak Aditya and Prof. Mainack Mondal

Worked in Google Cloud's Metering team. Developed an efficient pipeline to compute and export various storage consumption metrics of Google Cloud customers. The pipeline reduced the delay involved in metric computation by 73% (from 30 hours to 8 hours). I also received a pre-placement offer from Google for the full-time software engineer role.

Title: AuditLP: A Comprehensive Dataset and Benchmark for Studying Social Biases Encoded in WikidataDeveloped a dataset to benchmark the impact of social biases encoded in the Wikidata Knowledge Graph on the downstream ML models trained for the task of link prediction.Advisor: Prof. Animesh Mukherjee