I work on training coding agents for Copilot, focusing on improving its coverage metrics so that developers can delegate real work to the agent and trust the result.

A single profile rarely captures everything a person cares about. At Amazon I developed ASPEX, a framework that builds personas around the distinct aspects of a user's preferences. It induces a hierarchical taxonomy of aspects, assigns users to them softly, and refines persona prompts with OPRO. I evaluated it on eight personalization tasks across Amazon data and three public datasets, covering more than 300K users, with persona construction kept strictly separate in time from evaluation.

ASPEX outperformed contemporary persona-generation baselines, improving PR-AUC by a relative 24.2% and ROC-AUC by 15.8% on engagement prediction, and raising the recommendation follow-up rate by 18.9%. I also applied the same aspect-grounded idea as a verification step for production LLM inference on AWS Bedrock with Claude models, cutting end-to-end latency by 20.3%. The paper is submitted to the EACL 2027 Industry Track, and the project earned second place in the internal hackathon for applied scientist interns.

I worked on a tool-use agent built on Gemini 3.5 Pro, and on a question that kept surfacing: how well do frontier agents hold up when the world stops looking like their benchmarks? I built VISTA-Bench, an adversarial evaluation framework that generates diverse test environments through a symmetric adversarial protocol. It spans more than 300 synthetic mobile environments, tracks compositional generalization, trajectory consistency and recovery from multi-step errors, and cut evaluation overhead by 35%.

The results were sobering: agents scoring above 90% on standard benchmarks lost up to 42% of their success rate under distribution shift. We also found a surprise, that agents generalized 22% better out of distribution in low-resource languages than in high-resource ones.

I work on EDSim, an LLM-agent emergency department simulator in which GPT-4o-mini-driven patient, triage nurse, bedside nurse and physician agents run the full care journey. We are extending it toward an ICML 2027 submission. I check simulated wait times against historical data from a Level-1 trauma centre across five CTAS acuity levels, and the baseline outputs match the historical wait-time distributions.

The simulator also lets us ask what-if questions. In surge experiments over 326 simulated patient visits, a 60% increase in arrival rate raised mean length of stay by 75.4%, from 552 to 968 minutes. In a replica of a 79-bed, 15-physician department, arrival-to-assessment times rose 74.7% and post-disposition boarding rose 100.7% under stress, which shows where the bottlenecks form.

My summer at IIIT Bangalore focused on speech recognition for languages with very little labeled data. Using matching networks and self-supervised learning, and building on wav2vec 2.0 with alignment-aware cross-lingual transfer, I reduced phoneme error rates by 28% on code-switched speech. I also analysed phonetic confusion across 15 language pairs and showed that self-supervised pretraining can extend speech technology to 11 low-resource languages, even with fewer than 100 hours of labeled audio.

My thesis explored intelligent market analysis for trading systems, where my interests in mathematics, quantitative finance and machine learning meet. The resulting paper, Enhanced Trading Systems with Intelligent Market Analysis, was accepted to the main track of the 4th International Conference on Artificial Intelligence and Applications (ICAIA).

I assisted 56 students in this course, designing the course and lab material, leading hands-on tutorials on implementing machine learning algorithms, and offering one-on-one guidance on projects and debugging.

I guided junior students through robotics competition projects and ran workshops on robot programming. Our team finished as runner-up for PESU at the World Robotics Championship in Noida.