Curriculum Vitae

B.S. Mathematics and Computer Science

GPA 4.0

Reconciling Set-Valued Policy & Dead-End Discovery in Healthcare Reinforcement Learning: An Empirical Analysis

Gumbel-Based Active Sparse Mobile Crowd Sensing with Time Series Transformer

Patched Forecasting with Gumbel-Based Selector for Sparse Mobile Crowd Sensing

Ensemble Learning with Early Fusion of Kernel-Transformed and Classical Electrocardiogram Features for Chagas Disease Detection

Softening Overly Demanding Requirements in Recommendation System

Evaluating Digital Twins for Type 1 Diabetes by Decision Quality

CCQ: A Multi-State Child Care Quality Dataset to Support AI for Children’s Health Research

Physiological Identifiability of Type 1 Diabetes Digital Twins Under Behavioral Heterogeneity

Hierarchical Timeline Generation and Visualization for Child Care Quality Monitoring Reports

Type 1 Diabetes Digital-Twin Benchmark

• Proposed an evaluation protocol that scores a Type 1 diabetes digital twin by decision quality — whether its predicted glucose trajectories rank candidate insulin treatments as the real patient would (Spearman correlation, regret) — rather than by trajectory prediction accuracy (RMSE, MARD) alone.
• Built the benchmark on 30 UVA/Padova (simglucose) virtual patients under 21 basal-bolus treatments, comparing MCMC, simulation-based inference, and Kalman filter twins with linear, neural, and prior baselines.
• Showed the two disagree: a population-prior twin with the worst RMSE (63.6 mg/dL) ranked treatments second best (0.766 Spearman), while a neural twin with half the RMSE (32.9 mg/dL) ranked at chance (0.049), verified with Friedman and Holm-corrected Wilcoxon tests; first-author manuscript under review.

Child Care Quality (CCQ) Dataset

• Built CCQ, a de-identified dataset of 64,479 child care providers (29,152 quality-rated) across 12 U.S. states, collected from state QRIS portals and released as row-aligned text and preprocessed tabular versions on Hugging Face; first-author manuscript under review.
• Developed an LLM-based curation pipeline that adapts a hand-built Georgia reference to anonymize and clean the other states with open-source Qwen3 agents, verified against expert-curated ground truth.
• Designed within-state and leave-one-state-out benchmarks over tabular models (LR, RF, XGBoost, TabNN, TabPFN) and language models (ModernBERT; Qwen3-4B with classification head, RAG, LoRA), scored by balanced accuracy, quadratic weighted kappa, and Elo; tree models won within-state, zero-shot transfer was near chance, and modest target-state supervision recovered most within-state performance.
• Developed a causal-discovery graph (FPM + PC) and a counterfactual framework (DiCE) to prune, identify, and rank key determinants of high-quality child care in Georgia across 245 provider attributes.

Elk Preservation Application

• Architected and deployed the backend for a PERC elk-conservation web app: a PostgreSQL image-metadata database and cloud object storage (Google Cloud Run, Scheduler, Secret Manager, Cloud SQL, Storage), integrated with a Firebase-hosted React frontend for rancher wildlife-photo uploads.
• Fine-tuned a ViT classification model leveraging pretrained YOLOv9 object detection to estimate elk counts from images, supporting fair compensation for elk-related land damages by environmental agencies.

Healthcare Reinforcement Learning Research

• Conducted an empirical study on the consistency of Set-Valued Policies and Dead-End Discovery frameworks for clinician-in-the-loop sepsis treatment on MIMIC-III, and designed a partial ordering over recommended actions that gives clinicians actionable treatment recommendations.
• First-author paper accepted to the Pacific Symposium on Biocomputing (PSB '27) for publication in the proceedings and a poster presentation.

Sparse Mobile Crowd Sensing

• Developed a Learned Gumbel-Based Active Sparse Mobile Crowd Sensing framework: an encoder–decoder Time Series Transformer plus a custom sensor-selection layer that samples Gumbel noise from learnable weights, reducing reconstruction error of missing data by up to 28% over standard Time Series Transformers.
• Extended the framework with a patched variant that gives each patch of a sensing cycle its own selector layer, testing whether intra-cycle structure (e.g. morning, afternoon, night) improves selection.
• Benchmarked both on Urban Air (437 stations, hourly) and SensorScope St-Bernard (31 sensors, two-minute); found the patched variant improves only in isolated settings, likely from lost cross-patch attention.
• Published a first-author full paper and a follow-on extended abstract at IPCCC '25.

ECG Classification Research

• Developed an ensemble learning framework via AutoGluon, ECG-FM, FFT, and Wavelet Transform with early fusion for Chagas disease detection from 12-lead electrocardiograms.
• Competed as team GAIN-ECG and placed 39th in the 2025 George B. Moody PhysioNet Challenge.
• Published and presented a first-author paper at CinC '25.

Fairness in Recommender Systems

• Reduced item under-recommendation bias while cutting training costs in Bayesian Personalized Ranking by replacing an adversarial debiasing network with an autoencoder plus ranking post-processing.
• Published as a co-author at ISCAIS '23; all four authors contributed equally.

Math & CS Tutor (Part-time)

• Tutored undergraduates across the introductory calculus sequence (MATH 111 Calculus I, MATH 112 Calculus II) and the computer science core (CS 170/171 Introduction to Computer Science I–II, CS 253 Data Structures and Algorithms).

Languages

Python, Java, C, MATLAB, SQL

ML

PyTorch, NumPy, Pandas, SciPy, scikit-learn, XGBoost, Hugging Face Transformers, PEFT/LoRA, AutoGluon, TabNet, TabPFN, SHAP

Tools

Git, SLURM, Google Cloud (Run, Cloud SQL, Storage), Firebase, PostgreSQL, React, Playwright