Atri V. Sharma

Hi! I am Atri Vivek Sharma, a PhD student at Imperial College London, supervised by Prof. Alessio Lomuscio at the SAIL Lab. My research interests lie broadly in the safety and reliability of machine learning systems, with a particular focus on robustness, evaluation, and interpretability. I am especially interested in understanding how modern AI systems behave under adversarial, distributional, and interactive settings, and in developing methods for systematically identifying and mitigating failure modes.

My current work focuses on the robustness of large language models and AI agents, including the use of intent-preserving user simulation to evaluate multi-turn systems under challenging interactions. Furthermore, I have studied semantically equivalent adversarial attacks for eliciting intrinsic hallucinations in retrieval-augmented generation systems. Previously, I have worked on training methods for improving the adversarial robustness of XGBoost.

Alongside my PhD, I work at Safe Intelligence on synthetic data generation, adversarial testing, and specification-driven evaluations for AI agents in the spec27 product.

Before my PhD, I worked at Pangaea Data, a London-based startup applying AI in healthcare, where I led the development of an NLP-driven pipeline for identifying at-risk and undiagnosed patients. This experience shaped my interest in building reliable and interpretable machine learning systems, particularly in settings where model outputs directly inform high-stakes expert decision-making.

Outside of research, I love traveling, hiking and long-distance running!