Résumé

Updated September 2026

Education

EXPECTEDMAY 2028

The Pennsylvania State University

B.S. Computer Science, Minors in Physics and AI · Schreyer Honors College · GPA 3.71 / 4.00 · University Park, PA

SEP 2026– NOW

Do Residual-Stream Probes Survive Degrading Chain-of-Thought Legibility?

Supervised Program for Alignment Research (SPAR) · Mentee Researcher · mentor: Marios Tsatsos

To test whether activation probes still work when chain-of-thought turns illegible, sampled 1,816 Qwen3-4B traces on 400 borderline prompts (XSTest, OR-Bench) the model both complies with and refuses, so a probe must read the reasoning, not the prompt
Labelled comply/refuse with two LLM judges (88% agreement, κ = 0.76), since keyword checks misread soft refusals

AUG 2026– SEP 2026

Independent · PyTorch, jacobian-lens · Qwen3, Pythia, GPT-2, Gemma-2

Traced the per-token offset that buries the J-lens's readout of a model's activations to the model's own word frequency (Spearman 0.48); subtracting it, the obvious fix, makes readouts 3.5–7× worse
Z-scoring took secret-word elicitation from zero at every layer to the only working method on Qwen3-1.7B (pre-registered)

JUL 2026

Independent · PyTorch · LessWrong

Explained why one small transformer in a published memorization challenge could barely learn: its ReLU units died in training and cut off the gradient, a training failure, not a capacity limit
Confirmed it with one controlled experiment (a bias after the hidden layer restored full capacity, 1024 facts; before it, none, 488) and hand-built weights storing 672 facts; overturned the authors' explanation

JUL 2026

Independent · PyTorch, TransformerLens · LessWrong

Trained a 6-layer character-level transformer on Nietzsche's prose and located the one attention head that copies earlier text forward
Ruled out quote and parenthesis matching; the head has half a bracket-matching mechanism: what to write, not where to look

OCT 2025– AUG 2026

Penn State Multi-Campus REU · advised by Prof. Asif ud-Doula · Python

To measure the error from NICER's equal-glow assumption, simulated X-ray photons scattering through a neutron star's atmosphere
Validated it against an exact analytic solution (reduced χ² = 0.70) and published waveforms (0.11% peak error)
Showed the error shifts between two observables; first-author paper under review at Astrophysics and Space Science

DEC 2025– APR 2026

Nittany AI Advance Internship Program

Application Specialist · Team of 5 · Client: Penn State Office of the Physical Plant · State College, PA

Built a tool that pulls records out of scanned documents (OCR, LLM, regex); processed 1,000 client documents, with human review
Raised accuracy up to 10% by merging LLM and regex results and benchmarking Azure models to pick GPT-5; 5× faster in parallel

FEB 2026– NOW

CrashAI — Crash-Risk Analysis Platform

Founder & Technical Lead · Team of 3 · 1st of 40 teams, Nittany AI Challenge 2026 · Scoped with PennDOT

Built a crash-severity model (XGBoost, 81 road and driver features) with per-site explanations and what-if simulation; led a team building a 33-tool agent, retrieval over engineering manuals, and an eval harness

Machine Learning Python, PyTorch, HuggingFace Transformers, TransformerLens, NumPy, scikit-learn, Monte Carlo simulation, Git

Mechanistic Interpretability Circuits, OV/QK analysis, activation and path patching, direct logit attribution, Jacobian and logit lenses