What I’m Working On

I’m currently working on interpretability evaluation through the PRISM AI Safety Research Fellowship, using drug toxicity biology as a controlled testbed to evaluate whether interpretability tools such as linear probes and chain-of-thought faithfulness metrics can reliably identify how and why LLMs succeed or fail at causal reasoning. The broader question I care about is whether we can trust the methods we use to understand models in the first place.

Why This Work

We have built models that perform astonishingly well despite our incomplete understanding of how they work internally. This disconnect between capability and understanding became deeply concerning to me, especially given that these systems are already deployed at scale while we are still learning about their failure modes as we go. It feels like the safety community is scrambling to catch up.

My initial work in AI safety began with evaluations, examining model outputs, but over time I found myself less interested in simply observing what models do and more curious about understanding the internal mechanisms driving their behavior. What draws me now is the question of how to study that more rigorously, borrowing statistical and causal frameworks from other fields. Beyond my current fellowship, I’m curious about studying how representations form during training and using mathematical frameworks borrowed from physics to ask why neural networks organize information the way they do.

What I Bring

LLM Evaluation Experience

UVA Capstone · Deloitte Anthropic Alliance · 2024–2025

Led bias detection evaluation comparing LLM-as-a-Judge frameworks, processing 3,187 test cases across three complementary datasets. Discovered that Claude 3.5 Sonnet systematically failed on counter-stereotypical pronoun examples, revealing memorized gender associations over genuine syntactic understanding. The project won the program’s “Most Innovative Analytical Solution” award and was published in IEEE SIEDS 2025.

Technical Leadership and Mentorship

Emerson Automation Solutions · 2021–2022

Proposed and took on dual role to address unclear requirements blocking a team of junior engineers. Wrote detailed user stories, mentored developers, and improved team predictability while continuing to ship development work.

Cross-Layer Systems Debugging

Emerson Automation Solutions · 2019–2021

Led cross-team investigation of a critical cross-layer failure blocking a major product release. Traced a startup error loop in a Linux process interacting with a parallel runtime OS inside a hypervisor VM, a subtle interaction traditional debugging had missed entirely.

Strong Technical Foundation

5 years software engineering experience in industrial automation at GE and Emerson. MS Data Science (UVA, 2025) with coursework in Deep Learning, Bayesian ML, GPU Architectures, and Decoding LLMs. BS Electrical & Computer Engineering with minors in Mathematics and Computer Science. Proficient in Python, C/C++, PyTorch.

Completed Programs

  • BlueDot Impact Technical AI Safety course (June 2026)
  • BlueDot Impact AGI Strategy course (April 2026)
  • CEA High-Impact Career Pivot bootcamp (April 2026)