Built and productionised LLM systems over unstructured healthcare and insurance data.

• Built an LLM chatbot answering insurance agents' plain-English questions, from single-policy lookups to geographic aggregates, by compiling each into a dynamically generated Python retrieval pipeline. Increased the client calls each agent could handle by 25%.
• Engineered the data-acquisition backend powering the chatbot: a scraping-and-parsing pipeline that structured 3,000+ unstructured insurance policies into a queryable knowledge base.
• Architected an LLM-driven insight-mining pipeline over 100K+ unstructured healthcare documents, combining clustering with RAG-based entity extraction across 10K+ physician notes to cut retrieval latency by 40% and inform safety reviews and marketing strategy.

Part of the Microsoft Search, Assistance and Intelligence team.

• Developed voice assistant features for Microsoft Search, improving enterprise search accessibility and daily user engagement.
• Built RESTful APIs, frontend components and YAML conversational datasets alongside product and design teams, improving intent recognition and accelerating delivery of production features.

Undergraduate research: Transformer-based Models for Supervised Monocular Depth Estimation, published at ICICCSP 2022

Formal v/s Contextual Cues for Grammatical Gender Prediction

• Trained vector representations from the contextual and formal properties of words and compared their usability in a noun gender prediction task
• Used gender-debiasing methods to isolate the gender information in a word vector, showing that the human reliance on semantic over phonological cues shows up in language models too

Diachronic Distributed Word Representations as Models of Mental Lexicon

• Used a temporally sliced corpus to learn distributed word representations of child speech and child-directed speech under a curriculum-learning setting
• Demonstrated the ease of usage and effectiveness of diachronic distributed word representations in modelling lexical development, presented at the ACL 2022 Student Research Workshop