About
I study Computer Science and Mathematics in the Honors College at The University of Southern Mississippi. My research asks what language models actually do when we evaluate them: whether chain-of-thought explanations are faithful, whether benchmark scores measure what they claim to, and what internal features a model relies on.
Two lines of work: a behavioral audit of unfaithful chain-of-thought detection (arXiv preprint, under review at ACL Rolling Review) with Dr. Nick Rahimi through the CRA UR2PhD program, and a mechanistic audit of insider-threat language models with sparse autoencoders and causal interventions (manuscript in preparation) with Dr. Haiyan Tian. With Dr. Gregg Jacobs at the Institute for Advanced Analytics and Society, I analyze the errors of NOAA's RTOFS ocean forecasts against Argo floats.
Break Through Tech AI Fellow (Cornell Tech, 2026); ACM SIGHPC Immersion Program at SC25, one of 15 students selected nationally; Honors Keystone Scholar.
Research interests
Selected work
Behavioral detectors of unfaithful reasoning, evaluated against FaithCoT-Bench's human labels and stratified by answer correctness; linear probes on model activations where behavioral signals fail.
Started as an unsupervised anomaly-detection pipeline on CMU CERT data (Featuretools DFS, PCA, Gaussian mixtures); now a mechanistic audit of a QLoRA-adapted language model with sparse autoencoders and causal interventions.
Global analysis of NOAA RTOFS hurricane-heat-potential and 26 °C-isotherm-depth errors against Argo floats, with TEOS-10 profile processing, gradient-boosted corrections as the probe, and noise-column, shuffle, and forward-holdout tests.
Historical replay prototype for machine-learning-corrected mixed layer depth in the Gulf of Mexico, combining RTOFS model fields with nearby in-situ observations and provenance-aware confidence output.
End-to-end German eBay Motors NER pipeline that extracts structured product attributes from short listing titles.
Qiskit simulator for quantum walks that maps a Galton board to reusable n-layer circuits and compares quantum, classical, and biased distributions.
FastAPI-backed South Africa weather and air-quality service with gridded fields, point visualizations, GeoJSON industrial layers, and NumPy data pipelines.
CNN-based wildfire spread predictor paired with a React Native map interface, Firebase notifications, and FastAPI model serving on AWS EC2.
First-time COMAP MCM project on 2028 Olympic medal prediction. The team tested linear regression, ARIMAX, Bayesian and grey models, then chose Random Forest regression for the final pipeline. COMAP featured the team's story.
Presentations and talks
"Continuous Time Markov Chains for Insider Threat Behavior Modeling" (with H. Tian).
Lightning talk at Student HQ as part of the ACM SIGHPC Immersion Program, after an OLCF HPC/AI crash course. St. Louis.
"Principal Component Analysis for Data Compression, Noise Reduction, and Anomaly Detection: A Case Study on Insider Threat Detection" (with H. Tian), undergraduate research session.
AMS talk listingFellowships, awards, programs
One of 15 students selected nationally.
$2,000 research grant funding the Senior Honors Thesis.
$2,000 scholarship for faculty-led mathematics research.
23rd on the public leaderboard.
Successful Participant; team story featured by COMAP.
CompuCell3D Biological Modeling summer school and hackathon (Indiana University); WISER & Womanium Quantum summer program.
Additional experience
Built the data-ingestion pipeline and the review-and-publish controls for a NOAA-facing ocean data platform; shipped to production.
Encoded legislative rules and scoring metrics for Gulf Coast Restoration Fund projects into databases for optimization-based allocation.
Designed a FastAPI service serving gridded temperature, wind, and PM2.5 fields for South Africa.