I'm a master's student in Statistics (University of Central Florida, expected December 2026). My thesis develops Bayesian and frequentist uncertainty-quantification methods for Positronium Lifetime Imaging, an emerging extension of PET imaging, advised by Prof. Hsin-Hsiung Huang.
Most of my work comes back to one question: does a reported confidence deserve to be believed? That started with estimators in medical imaging and now extends to AI systems — whether the number on the dashboard is measuring what it appears to measure.
- Languages: Python, R, SQL
- Methods: Bayesian Inference & MCMC, Uncertainty Quantification, Calibration & Coverage, Regression (linear, logistic, beta, ridge, lasso), GLM, Design of Experiments, Machine Learning
- AI systems: retrieval-augmented generation, agentic tool-use workflows, LLM evaluation
- Tools: PyTorch, statsmodels, scikit-learn, spatial data (sf, terra), Claude, LaTeX, Tableau
- rag-eval-audit — An agentic RAG system over public banking-regulatory documents, plus an evaluation layer that audits both the system and the judge grading it. It cited rescinded guidance as current authority on 83% of runs — and the standard groundedness metric scored every one of those failures 5/5.
- statistical-ml-implementations — ML methods built from the math up: CNNs, SVDD with directional-data kernels, an autoencoder for anomaly detection, and transformers.
- fl-conservation-regression — Beta regression on what drives land conservation across Florida's 67 counties, built end-to-end from raw spatial data (FGDL, NLCD, Census) to a written executive summary.
- paper-helicopter-doe — A hands-on design-of-experiments study (factorial and response-surface designs) optimizing paper-helicopter flight time from data I collected.
- 📫 katrina.a.stephenson@gmail.com
- 🎓 Member, American Statistical Association