Simple Averaging vs. Learned Stacking in K-Fold Ensembles
Simple averaging beats learned stacking as architectural diversity grows. R² 0.89 on blind wells, +7.16 pp over the best single model.
Software engineer turning ambitious ideas into working solutions. Currently pursuing an MS in Computer Science at Northeastern while teaching machines to learn and students to think.
Index 01 · The Premise
I build at the intersection of shipping software and pushing research. After converting an Oracle internship into a full-time role on Oracle Health, and earning a Star Performer Award along the way, I went deep on the questions that wouldn't let go.
Now I split my days between graduate AI/ML coursework, a research lab, two classrooms I teach, and an advisory board. The throughline: turn ambitious ideas into things that actually work, with a researcher's rigor and an engineer's "but does it ship."
Academia · Industry · Research · Teaching
Index 03 · The Journey
Most timelines stack life into a single column. Mine doesn't fit. Scroll across the lanes, Academia, Industry, Research, Teaching, Service, and watch them overlap.
The convergence
In Fall 2025, I held four roles at once.
Key Milestones
Simple averaging beats learned stacking as architectural diversity grows. R² 0.89 on blind wells, +7.16 pp over the best single model.
VGG19 + custom CNN with Bayesian hyperparameter optimization. 96% on Montgomery, 86% on Shenzhen.
View publication →Index 04 · Selected Work
Real models, real metrics. From multi-agent medical AI to pore-pressure ensembles and medical imaging.
A 10-agent medical consultation platform. Safety screening, RAG over clinical guidelines, and reasoning, under 2 seconds end to end.
A 5-architecture meta-ensemble for pore-pressure prediction across 11 formations, with physics-informed features and SHAP and LIME.
U-Net segmentation of brain tumors from multi-modal MRI (T1, T1CE, T2, FLAIR), reaching clinical-grade overlap.
A four-phase concurrent Python pipeline unifying Semantic Scholar, OpenAlex, and GitHub to find and catalog every paper citing NDIF.
Predicts loan default risk using ensemble methods with feature engineering across borrower profiles.
A web app that detects and identifies multiple celebrity faces in an image, served through Flask.
A curated slice of the work. The full set spans AI/ML, full-stack, research, and hackathons.
See all work →