Principal Applied Scientist · Microsoft
I'm a Principal Applied Scientist at Microsoft, working on AI-agent evaluation and conversation analysis. Previously, I worked on production NLP at Qualtrics and model compression at Amazon Alexa AI.
My experience includes leading NLP projects, mentoring scientists, developing model-distillation code, and fine-tuning language models. My published research covers multilingual model efficiency, LLM evaluation, and matrix approximation.
01 / Experience
Applied-science roles at Microsoft, Qualtrics, and Amazon, spanning agent evaluation, production NLP, and efficient language models.
December 2025 – present
Principal Applied Scientist
AI-agent evaluation, conversation analysis, and end-to-end agent quality.
April 2021 – December 2025
Staff ML Applied Scientist
Production NLP, efficient multilingual models, and LLM adaptation, with technical leadership and scientist mentoring.
May 2018 – April 2021
ML Applied Scientist
Knowledge distillation, compressed language models, architecture search, and speech-model research.
02 / Selected publications
2026 · EACL Industry
Updating one language should not require starting again with every language. Our coauthored study evaluates language-specific model merging across three tasks. It reports up to 50% less initial training time and more than 60% lower training cost for individual-language updates, while maintaining quality parity in the evaluated settings.
Read the paper2023 · IAAI / AAAI
How do you bring multilingual NLP into production without multiplying deployment overhead? Our work combines cross-lingual, multi-task modeling with distillation. The published study covers classification in 12 languages and reports estimated hardware-cost savings of 44% compared with the original full-size model. Those are study-specific estimates, not a general guarantee.
Read the paper2023 · EMNLP · GEM Workshop
Evaluating language models on production data can be costly when every assessment requires fresh human labels. This coauthored study explores ensemble disagreement as a proxy for model error on keyphrase extraction across studied languages and domains. It provides evidence for a lower-cost evaluation signal, not a universal replacement for human judgment.
Read the paper03 / Research foundations
Ph.D. Computing, University of Utah, 2017
My dissertation, Data Analysis and Visualization using Basis Selection for Matrix Approximation, studied compact representations of large datasets. The work applied matrix sketching and basis selection to kernel learning, simulation science, and visualization.
Explore the research04 / Contact
If you'd like to connect, reach out on LinkedIn.
Connect on LinkedIn