My work revolves around the question:
How can agentic systems reliably act on human intent?
At Microsoft, I build agentic loops and evaluation harnesses for enterprise AI scenarios. Currently, I am interested in improving reliability and memory management in long-horizon agentic tasks.
Previously, I obtained my Ph.D. in Computer and Information Technology from Purdue University. My research focused on social reasoning and reliability of LLMs on manipulative discourse, where intent is ambiguous due to the use of deceptive language. During this period, I completed two research internships at Microsoft Research (2022, 2023), working on leveraging LLMs for template discovery from streaming document collections.
When I am not working, you will find me with a mystery novel, on a trail somewhere, or painting. Always up for a chat — let's talk!
Research Lens
-
Intent Classification
Parsing user goals under ambiguity and multi-intent framing.
Related work
- Reasoning over grooming & coercive discourse (Ph.D. thesis, Purdue 2025)
- Grooming risk estimation with fuzzy theory (NAFIPS 2024)
- Fuzzy classification of multi-intent utterances (NAFIPS 2022)
-
Reliability
Making model outputs trustworthy enough to act on.
Related work
- The Reliability Paradox: shortcut learning undermines calibration (HICSS 2025)
- Learning Shortcuts: the misleading promise of NLU (AIDPS 2025)
- Calibration error estimation via fuzzy binning (NAFIPS 2023, honorable mention)
-
Responsible AI
Surfacing bias, shortcuts, and failure modes that shape safer LM behavior.
Related work
- Hire Me or Not? Occupation attributes shift LM decisions (COLING 2025)
- Fuzzy evaluation of sentence encoders on risk classification (NAFIPS 2024, outstanding student paper)
- Model choices influence attributive word associations (WI-IAT 2020)
Selected Publications
- COLING 2025 2025
Hire Me or Not? Examining Language Model Behavior with Occupation Attributes
Occupation attributes systematically shift decisions made by the LLMs that increasingly drive agent workflows — exposing behavior that matters for responsible deployment.
- HICSS 2025 2025
The Reliability Paradox: Shortcut Learning & Language Model Calibration
Calibration alone can hide shortcut reasoning inside the LLMs that power agents — reliability needs deeper signals before we trust agent outputs.
- NAACL 2021 2021
Low Anisotropy Sense Retrofitting (LASeR)
Reshaping embedding geometry so LLMs preserve richer word-sense information — cleaner representations for the models that reason underneath agents.
News
- Mar 2026Joined Microsoft as a Senior Applied Scientist, working on agentic systems for customer experience applications.
- Feb 2026Invited panelist at M3AAWG on AI-generated content, online harms, and trustworthy systems.
- Aug 2025Defended my Ph.D. dissertation at Purdue University on reasoning capabilities of language models for detecting grooming and coercive discourse.
- Mar 2025Received a Graduate Student Travel Grant from the Purdue Polytechnic Institute.
- Jan 2025Delivered an ignite talk, “Bridging the Gap: Advancing AI for Detecting Covert Online Harms,” in the Digital and Social Media track at HICSS-58.
- Dec 2024Our paper, “Examining Language Model’s Behavior with Occupation Attributes,” was accepted to COLING 2025. Paper.
- May 2024Received an Outstanding Paper Award at NAFIPS 2024 for “Evaluating Language Models on Grooming Risk Estimation Using Fuzzy Theory.”
- Apr 2024Awarded Third Place at the 2024 Purdue Polytechnic RIA Student Poster Session for “Detecting Online Grooming Risk Using Transformers.”