Research and applied work, from wearable-sensing systems to coursework and internships.
A wearable broad-band optical spectroscopy system for noninvasive lactate estimation during exercise. It captures broadband optical signals across lactate-sensitive bands and applies a two-stage hierarchical regression — a PLS model for the per-session trend and an ElasticNet model for the absolute level. Validated in vitro and in an in-vivo cycling study (11 participants, VO₂max protocols), reaching R² ≈ 0.62. Paper under review.
The full hardware-and-software stack for the Lumos multi-wavelength wearable spectroscopy device: Arduino firmware across board revisions (9–11 LEDs, up to 940 nm), Python analysis tooling, and a companion Swift iOS app.
An open-source Python package for storing, managing, analyzing, and visualizing multi-wavelength PPG measurements from the Lumos wearable. Each Adub object holds one subject's measurement, with dataset storage, export, and quick-view utilities.
A modular, open-source platform that acquires, visualizes, and optimizes spectroscopic biomarker profiles. I built the data structures and server architecture for multi-channel wavelength time-series from custom wearable biosensors, plus a spectral-attention model and a curated biomarker database that improves PPG sensor performance.
A multi-wavelength optical sensing framework for cuffless, calibration-free blood-pressure estimation from wearable PPG. I contributed to the modeling and analysis of the multi-channel optical signals. Published at ICASSP 2025.
Authored a comprehensive review of fetal-brain-MRI motion-correction methods — spanning CNNs, LSTMs, Transformers, GANs, and diffusion models — and their implications for prenatal imaging.
A unified, full-spectrum benchmark for OOD detection. I helped analyze 100+ papers and benchmark ~40 methods on ImageNet-1K, and built an online leaderboard and a lightweight OOD evaluator. Accepted at the NeurIPS 2023 DistShift workshop.
Explored using LLMs as a memory architecture for autonomous agent control in a simulated MiniGrid world. I generated 250K GPT-3.5 tokens, engineered prompts for decision-making and experience summarization, and designed a compact representation of observations, status, and past actions within a limited context window.
Studied attack classes that degrade neural networks and adversarial-training schemes (student–teacher, teacher-free, self-distillation), then designed an efficient adversarial- training method based on knowledge distillation.
Proposed a directed-graph clustering algorithm based on marginal propagation, framing community detection on the Email-EU-Core dataset as a semi-supervised learning task.
Investigated automated atomic-model building from cryo-EM density maps for structural biology.
An iOS app showing real-time reviews and menus for Duke's West Campus cafeterias. Built with SwiftUI and a Vapor server on AWS EC2 (SQLite), integrating MapKit and Duke's Streamer API, with App Clips launched via QR codes and NFC tags.
Analyzed 122,411 elite-match gameplay snapshots with neural networks and PCA to correlate equipment usage with round outcomes and surface game-turning strategies.
A full-stack rental-management platform: a MySQL data model with ER diagrams, an object-oriented PHP backend on MAMP, and an HTML/CSS/JS interface with token-based registration links.
A recommendation engine over 1B+ IMDb entries (2000–2020) combining collaborative filtering (ALS), Word2Vec NLP, and logistic regression (RMSE 1.15), with TF-IDF similarity across 2M+ words.
Designed a tritanope accessibility algorithm that optimizes web and app color layouts for WCAG 2.0 compliance, plus Python/VBA data-migration and triage scripts that improved testing efficiency by ~50%. Recognized with the Outstanding Intern Award.