Statistics and machine learning research, outside the resume line items.

2023 – 2024

Ongoing research, unpublished

Faculty Perspectives on Game-Based Learning

Dr. Michael Rugh · LIVE Lab, Texas A&M

Game-based learning research usually studies students. This project studies the other side, faculty attitudes toward using games in teaching. I worked in the LIVE (Learning Interactive Visualization Experience) lab through literature review, survey analysis, and Structural Equation Modeling (SEM) in STATA to identify what drives variance in faculty survey responses.

• Used SEM path diagrams to model relationships between latent, hard-to-measure concepts (e.g. attitudes toward pedagogy) via directly observable survey indicators
• Ran Confirmatory Factor Analysis (CFA) in STATA to test a hypothesized model against real survey data, rather than building the model from the data itself
• Collaborated with a five-person research team (Yanwen Chen, Emma Koh, John Blough Christian, Nishant Gandhe) through multiple rounds of model iteration and debugging

2023 – 2024

Literature-based publication; quantitative work not included in final paper

Kru Maritime History Archive

Dr. Megan Crutcher · Nautical Archaeology, Texas A&M

A five-person project reconstructing the history of the Kru, a seafaring ethnic group from present-day Liberia whose sailors were hired, rather than enslaved, by European traders for their navigational skill during the Atlantic slave trade era. The core work was literature review and transcription of interviews conducted with Kru elders; I took on the quantitative side, analyzing 1800s Portuguese shipping records.

• Learned R for the first time to analyze historical shipping-record data extracted from archival Portuguese sources
• Ran hypothesis tests (including a t-test comparing sailor age by origin) and built visualizations of sailor counts by origin and occupation, and geographic origin mapped to former Portuguese colonial territory
• The final published piece took a purely literature-based approach per the co-authoring team's direction, so this quantitative analysis was a personal skill-building exercise rather than part of the published output

2024 – 2025

Published in ASEE PEER (Work in Progress track), 2025

Perspectives of Junior Scholars: Calculus Learning Outcomes from an Educational Video Game

Dr. Michael Rugh · LIVE Lab, Texas A&M

A companion project to the faculty-attitudes study above, from the student side. This paper studies how middle schoolers actually experienced Variant: Limits, a video game used to teach calculus concepts (limits, trigonometry, geometry) at Texas A&M's Aggie STEM Summer Camp 2024, an audience the game wasn't originally built for.

• Helped design and facilitate the summer camp sessions where 30 middle schoolers played the game, then worked with the resulting interaction data as part of the LIVE Lab's broader effort maintaining its educational-game database
• The study analyzed 30 surveys and 6 interviews (181 coded text chunks) using thematic coding plus 1–7 Likert sentiment scoring across six themes
• Findings were mixed. Students rated the game's actual instruction (M=2.88) and controls (M=2.44) fairly negatively, while general opinion of educational video games stayed neutral-to-positive (M=4.67), pointing at execution, not the format itself, as the real problem
• Presented as Board #213 at the American Society for Engineering Education (ASEE), the first of my LIVE Lab work to reach a public academic venue

Spring 2025

Honors contract · Formal paper completed

Credit Score Prediction in the Modern Age

Honors contract with Dr. Scott Bruce · STAT 315, Texas A&M

The capstone project for STAT 315 (Computational Data Science), extended into an Honors contract requiring work beyond the standard course requirements. I helped lead a team building a Random Forest model to classify credit scores into Poor, Standard, and Good categories, then wrote a publication-ready paper documenting the methodology and findings — the deliverable that distinguished the Honors section from the standard course.

• Built and tuned a Random Forest classifier for three-class credit score prediction as the team's technical lead
• Completed an Honors contract requiring a publication-ready paper beyond the standard course deliverable, applying methods from STAT 315 in a full project workflow modeled on industry practice
• Applied course-wide methods (Python, SQL, bootstrapping, Docker, Git/GitHub) across the full project lifecycle

Spring 2026

Honors contract · Formal paper completed

Statistical Inference for Network Data

Honors contract with Dr. Chainarong Amornbunchornvej · STAT 415, Texas A&M

An Honors contract extending STAT 415 (Mathematical Statistics II) into an active research frontier: statistical inference for network, or graph, data. Where the standard course studies inference for independent samples, this project examines settings where observations are random networks drawn from an underlying distribution — a harder problem because the structure of the data itself is informative and observations are no longer independent.

• Developed the Erdős–Rényi (ER) model as a tractable starting point, deriving MLE and method-of-moments estimators for edge probability p, proving both unbiasedness and consistency, and building confidence intervals and Wald tests from the CLT
• Validated inferential tools via Monte Carlo power studies (B=1,000 replicates), confirming the tests perform well even at moderate network sizes (n ≈ 30)
• Extended to two-sample inference, testing whether two networks share the same edge probability — then applied the framework to the C. elegans neural connectome, comparing electrical vs. chemical synapse networks (n=279 nodes each) and confirming they are significantly different (Z=29.65, p≈3.5×10⁻¹⁹³)
• Discussed limitations of the ER model (independence assumption, single global p) and outlined more expressive alternatives including the stochastic block model, which connects directly to the latent-structure network research in machine learning

Jan – Apr 2026

Deployed as a live web app

Lung Cancer Detection from CT Scans

Team project with Pranav Gaddam, Kyle Wu

Most CT scan classification writeups report one final accuracy number and stop there. This project instead ran a controlled comparison of four modeling approaches under identical conditions, then isolated exactly what mattered for a small, imbalanced medical imaging dataset, the value of a domain-adapted starting point over generic pretrained weights. I built and compared all four models; teammates built the deployment app.

• Built and compared a PCA + Logistic Regression baseline, a fully fine-tuned Swin Transformer, and two transfer learning variants differing only in their starting checkpoint
• Best model (transfer learning from an already domain-adapted checkpoint) reached 91.2% test accuracy and 0.91 macro F1, a 41-point improvement over the classical baseline's 50.9%
• Isolated initialization as the key variable by holding training procedure identical between the two transfer learning models; the domain-adapted checkpoint beat fresh ImageNet weights by 7.6 points
• Addressed severe class imbalance, the baseline missed 93% of the most common cancer subtype, with inverse-frequency loss weighting

Spring 2026

Formal paper completed

Motor Symptom Prediction from Voice Measures in Parkinson's Disease

Team project

Parkinson's disease affects speech, and voice can be recorded remotely and repeatedly, which makes it a natural target for symptom monitoring outside a clinic. This project used the UCI Parkinson's Telemonitoring dataset, 5,875 observations from 42 patients, to test how well detailed voice measures predict motor symptom severity, and whether that relationship differs by sex. I led the multicollinearity diagnosis, variance inflation factor analysis, and model selection work.

• Diagnosed severe multicollinearity within voice predictor families (some VIFs exceeded 10^8) using pairwise correlations, an auxiliary regression, and generalized VIF, then reduced the predictor pool to resolve it
• Ran an all-subsets search across 17 candidate predictors comparing BIC and AIC selection; the final 11-predictor model found sex, jitter, shimmer, and nonlinear voice features all independently associated with motor symptom severity
• Found female patients have significantly lower average motor symptom severity than male patients at equivalent voice measurements (95% confidence interval entirely below zero)
• Built a separate predictive model and compared it against the interpretability model on out-of-sample error, trading 34 more predictors for a modest gain in prediction accuracy