EIGHT YEARS,FIVE LANES.

Four years of core computer science (data structures, algorithms, databases, operating systems). It is also where I wrote my first real research code and got hooked on building things that learn.

Self-driven research detecting tuberculosis from chest X-rays using convolutional networks (VGG19 plus a custom CNN) with Bayesian hyperparameter optimization. 96% accuracy on Montgomery, 86% on Shenzhen. Two years of work that became my first peer-reviewed paper.

Built a C# application for automated script analysis that improved developer productivity by 30%, and integrated a BDD testing framework (Cucumber, Serenity) into the CI/CD pipeline. Enough to earn the return offer.

I built and maintained clinical pharmacy software inside Oracle Health, the kind real pharmacists depend on. A multi-tier sorting algorithm cut clinical data access time by 60%. Print infrastructure I architected served 145,000+ endpoint devices with 40% faster load times through query optimization and caching. Backend work cut API response times by 55%. I resolved 6 critical CVEs in services handling PHI under HIPAA, and fixed 70+ cross-platform issues while modernizing legacy JavaScript. Along the way I mentored 3 engineers and worked across product, QA, and clinical teams. Stack: C#, C++, JavaScript, SQL.

My MS has been less about sitting in class and more about doing the work. Coursework went deep on Machine Learning, Deep Learning, NLP, Computer Vision, Advanced Perception, Scalable and Distributed Systems, Algorithms, and Generative AI, but most of my hours went into research and teaching outside of it. GPA 4.0/4.0.

TA for Machine Learning (CS 6140) and Deep Learning (CS 7150). I mentored 250+ graduate students, ran weekly coding sessions and office hours, and designed lab materials, problem sets, and evaluation rubrics that move students from theory to working implementation. Teaching is where I find my own gaps fastest.

I ran two lines of research at Khoury. First, GeoMeta: a 5-architecture meta-ensemble (CNN, DFNN, Transformer, Random Forest, XGBoost) for pore-pressure prediction across 11 geological formations, with physics-informed features (Eaton ratio, pressure gradients, normalized depth) and SHAP and LIME explainability. On 281,000+ samples from 21 wells it reached R squared of 0.92 on blind test wells, 27% better than the best single model, at under 50ms inference. Second, synthetic well-log generation: evaluating VAEs, GANs, and diffusion models to produce realistic tabular well logs under severe data scarcity, validated with a train-on-synthetic, test-on-real protocol. This research led to the IJCNN 2026 paper.

Early-stage exploration of autonomous drone systems for wildlife deterrence, before I moved fully into the GeoMeta research.

Liaison between the MS student body and Khoury leadership, turning student feedback into concrete suggestions on academic programs and curriculum.

NDIF, the National Deep Inference Fabric, gives researchers remote access to the internals of large models. As Technical Outreach Engineer I built hands-on tutorials on mechanistic interpretability using nnsight and nnterp, turning dense papers into reproducible, runnable code at a pace of one to two mini paper tutorials a month. I also led a full rebuild of NDIF's web presence, moving it from static HTML to an interactive Next.js, TypeScript, Tailwind, and Framer Motion site with a live, API-connected model status dashboard.