Research

My research focuses on consistency in multi-agent large language model (LLM) systems. I am a Ph.D. student in the KAI² Lab at UMBC, advised by Dr. Manas Gaur.

Current focus

Ongoing Ph.D. research

Consistency in multi-agent LLM systems

I approach this problem in three steps:

Define

What does “consistent” mean for a multi-agent system?

Consistency is usually defined for one model and one prompt. I'm working on a definition that fits settings where many agents interact.
Diagnose

Where does inconsistency come from?

I'm tracing inconsistent behavior back to its root causes, so that it can be understood and fixed rather than just observed.
Measure

How do we evaluate it?

I'm working toward metrics and benchmarks that measure consistency in multi-agent settings, so different systems can be compared.

Results from this project will be posted on my publications page as they appear.

Earlier work

My earlier research was in security, using ultrasound sensing on mobile and edge devices and analyzing malware source code.

Ultrasound-based security on edge devices

This work uses inaudible ultrasound signals on smartphones and other edge devices to verify that a human, rather than an AI agent, is present, and to watermark audio against voice scams.

EchoGuard: continuous ultrasound-based verification that a human, not an AI agent, is present (EdgeSP 2026 at ACM/IEEE SEC 2026, co-first author).
Ultrasound Watermark: real-time acoustic watermarking against voice scams on smartphones (EdgeSP 2026 at ACM/IEEE SEC 2026).

Understanding malware evolution

MASCOT is a manually reviewed dataset of 6,032 malware source-code samples. It is used to study malware from a software-engineering perspective and to trace how code reuse drives malware lineages.

MASCOT (arXiv preprint, 2025).

Methods & tools

Methods: Machine learning, signal processing, wireless and acoustic sensing, adversarial methods, malware analysis
Programming: Python, MATLAB, Kotlin, Java, Haskell, Linux, LaTeX
Tools: Android Studio, Adobe Audition