Experience

Applied Scientist · Jul 2026 – Present

Building foundation models of customer behavior — large generative sequence models — and applying them to counterfactual and causal estimation for large-scale decisions.

Applied Science Intern · May 2025 – Aug 2025

Built an end-to-end AI agent on AWS Bedrock that turned complex model outputs into actionable insights, combining LLM-powered reasoning with a retrieval-augmented grounding pipeline and an LLM-as-judge evaluation framework to ensure accuracy and robustness.

Developed a TMK-based self-model for an AI system for scientific modeling, enabling LLMs to generate metacognitive self-explanations grounded in an AI agent’s task, method, and knowledge structure.

Designed an iterative reasoning pipeline composing explanations across TMK components with traceable evidence; work resulted in a Springer book chapter.

Evaluated LLMs for the hard-constrained task of program synthesis by example (PBE) and designed techniques to improve robustness to prompt sensitivity.

Developed a neuro-symbolic pipeline to repair inaccuracies in LLM-generated code, solving over 30% of previously unsolved benchmark tasks across Python and Excel.

LLM-Guided Compositional Program Synthesis (arXiv:2503.15540, 2025)

Spearheaded the revamping of commenting functionality for a web-based document review workflow used by 100+ enterprise customers, including live synchronization of comments in multi-user environments.

Built an open-source customization plugin for Adobe Experience Manager, enabling enterprise users to customize the platform without legacy-code knowledge; received a Spot Award.

Filed a patent and improved F1 by 50% for Adobe Acrobat’s accessibility-tagging AI pipeline by developing an algorithm to incorporate and propagate human feedback.

Prepared Mudra, a systematically curated multimodal Hindi co-speech gesture dataset of 43.62 hours of videos with extracted poses, transcripts, and audio features.

Generated culture- and gender-influenced co-speech gestures for a conversing virtual agent, and conducted pilot studies and statistical analysis to confirm gender and culture influences. Used CNN and deep feedforward neural networks to extract and represent gender as a spectrum.

Advised by Dr. Jainendra Shukla (IIITD) and Dr. Aniket Bera (Purdue University).

Read the thesis

Estimated uncertainty when answering queries in the absence of a true causal graph. Compared structure-learning algorithms and evaluation metrics, and built a query-based ensemble to output the best algorithm for a given query on a dataset.

Demonstrated an 88.65% speedup over existing strategies for runtime verification of JavaScript programs.

Designed and implemented specification and instrumentation mechanisms to allow dynamic enforcement of properties, ensuring correct behavior and detecting deviations.

Worked as a research assistant on a study of digital journalism practices in India, interviewing journalists across newsrooms and supporting the study’s survey and analysis. The work was later published in Economic & Political Weekly.