The Research

RESEARCH THATHOLDS UP.

Two publications, two active workstreams, and a habit of checking whether the benchmark measures what everyone assumes it does.

2 publications · 2 active workstreams

01 · Publications

IJCNN 2026 (WCCI, Maastricht)

Accepted Mar 2026

Simple Averaging vs. Learned Stacking in K-Fold Ensembles

A simple question with a surprising answer: when you build a K-fold ensemble, should you learn how to combine models, or just average them? Using a leakage-free GroupKFold setup (no well appears in both training and validation, so out-of-fold R squared ranks models without ever touching ground truth), I show that simple averaging consistently beats a learned Ridge meta-learner, and the gap widens as you add more architecturally diverse models, because learned weighting discards diversity that matters under distribution shift. A Top-K ablation shows monotonic gains across 5 diverse architectures. The 4-model ensemble reaches R squared of 0.89 on 4 fully blind test wells, 7.16 points better than the best single model. Data: 188,631 training samples (17 wells) and 84,973 blind test samples (4 wells), Potwar Basin.

0.89

R² on 4 blind wells

+7.16 pts

Over best single model

188,631

Training samples

Preprint soon

Springer SSIC

Oct 2023

Detection of Tuberculosis from Chest X-Rays using CNN

CNN architectures (VGG19 and a custom network) for automated tuberculosis screening from chest X-rays, tuned with Bayesian hyperparameter optimization and validated on two public datasets. Two years of self-driven work that became my first peer-reviewed paper.

96%

Accuracy, Montgomery

86%

Accuracy, Shenzhen

Read the paper →

02 · Active research

Both workstreams ran under the Khoury GRA, May 2025 to Jun 2026.

GeoMeta, Pore-Pressure Prediction

A 5-architecture meta-ensemble (CNN, DFNN, Transformer, Random Forest, XGBoost) that resolves domain shift across 11 geological formations. Physics-informed features plus SHAP and LIME for interpretability. 281,000+ samples, 21 wells. R squared of 0.92 on blind wells, 27% over the best single model, under 50ms inference, with strict well-based splitting so there is no leakage.

Meta-ensemblePhysics-informedSHAP + LIMER² 0.92

Synthetic Well-Log Generation (GenAI)

Investigating VAEs, GANs, and diffusion models to generate realistic tabular well logs where data is scarce. A CVAE with a Conv1D and LSTM encoder, a dual-head decoder, and a physics-constrained loss, validated with a train-on-synthetic, test-on-real protocol on held-out wells.

CVAEGANsDiffusionTrain-synthetic, test-real

03 · Research projects

Privacy-Preserving Fall Detection via LoRA-Adapted Depth Estimation

2025 · Advised by Prof. Bruce Maxwell

Fall detection that never touches RGB at inference. I adapted Depth Anything V2 (a 25.4M-parameter foundation model) with LoRA, training just 2.32% of parameters, and lifted indoor depth accuracy (δ₁) from 0.36 to 0.82, a 128% relative jump, while cutting RMSE by 60%. A ConvLSTM then classifies falls from the depth sequence. I also found a systematic depth confound in the widely used UR-Fall benchmark (mean depth 0.448 for daily activity versus 0.593 for falls, p < 10⁻⁶⁶), and a temporal shuffle test showed a 0% accuracy drop, meaning models were ignoring time entirely and prior baselines were measuring the wrong thing.

0.82

δ₁ accuracy, up from 0.36

2.32%

Of parameters trained (LoRA)

−60%

RMSE

The most interesting result was negative.