ML, simulation, and optimization for real-world problems
Research project with Dr. Raman Arora comparing reinforcement learning alignment methods (PPO, GRPO, and DPO) for language models on verifiable reward tasks. Evaluating accuracy, training stability, and convergence using GSM8K math benchmarks.
Two-stage EfficientNet-B0 classifier for 8 tumor categories + glioma subtype analysis with Grad-CAM visualization. Includes a methodological audit quantifying evaluation contamination: patient-level leakage inflates benchmarks by +2.5 pp overall and +11.6 pp for specific classes. Clean rebuild with patient-aware splits achieves 96.4% on 4 trustworthy classes. Paper in progress: Quantifying Evaluation Contamination in Brain Tumor MRI Classification Benchmarks.
Advanced ML pipeline for differentiating tremor types (Parkinsonian vs. Essential Tremor) from wearable time-series data. Using XGBoost, Random Forests, and deep neural networks with dimensionality reduction and SMOTE for class balancing.
Built with the JHU Sports Analytics Research Group under Dr. Anton Dahbura. Probabilistic lineup optimization using Markov-style transitions and stochastic simulations. Developed BRP (Baserunner-Dependent Net Run Production), a custom metric for expected run contribution. 2025 JHU Design Day Digital Vanguard Award winner.
Co-founded EdTech startup as CTO. Built full-stack web infrastructure with automated scheduling, curriculum delivery, and performance tracking. Serving students from underserved communities.
Internal web app for Affordable Tutoring Solutions that generates Personal Study Material assignments for SAT students, tracks practice exam scores, and visualizes domain coverage per student. Features a worksheet catalog with subject/domain/difficulty filtering, per-student assignment history, a coverage heat map, and PDF-parsing pipelines for WellEd and ZipGrade score reports. Built as a single static page (React 18 + Babel in-browser) backed by Firestore.
Contributed to JHU's core simulation engine modeling stochastic disease spread across communities. Optimized NumPy-based pipelines and refined geographic modeling for improved accuracy.
Exact stochastic simulation of neural populations using the Gillespie algorithm. Models refractory periods, correlated variability via Cholesky decomposition, and network coupling. Includes parameter presets, Fano factor / CV statistics, and CSV data export.
LN and GLM encoding models predicting retinal neuron responses to visual stimuli. Features spike-triggered average estimation, model comparison metrics (AIC, log-likelihood), tuning curves, and data export for spike trains and model parameters.
Signal processing and ML pipeline classifying motor intent from synthetic ECoG data. Extracts frequency-band features (beta desynchronization, gamma activation) and trains a Random Forest classifier with calibrated probabilities across 4 movement classes.
Late-fusion classifier for neuron types using electrophysiology, morphology, and transcriptomics. Per-modality GradientBoosting classifiers feed a meta-classifier with modality dropout augmentation, achieving 90% fusion accuracy across 11 cell types.
Predicts neural connections from 3D morphology and cell-type features using gradient boosting on pairwise neuron features. Trained on synthetic connectome data following published cortical wiring rules, achieving 82% AUC on held-out networks.
Predicts Alzheimer's progression from clinical features (MMSE, hippocampal volume, APOE4, biomarkers) using gradient boosting for risk stratification and trajectory regression. Trained on ADNI-matched synthetic data with realistic biomarker correlations.