Work

Data Science Intern — St. Louis, MO

Python · LiDAR / point-cloud processing · ML pipelines · HPC · Gemini API · Claude Code

I worked on three distinct projects during my summer at WWT.

The main deliverable was a civil-surveying application that converts terrestrial laser scanning (TLS) LiDAR point clouds into engineering-grade CAD (DXF) models through an end-to-end ML pipeline. I extracted road and building shapes from large-scale scans using geometric and VLM-based methods, built a review UI for human correction, and consolidated the pipeline into one workflow projected to cut extraction time by 50%+, an estimated $12.5M in annual revenue for the client civil-surveying company. Scans ranged from a 78M-point public dataset (Toronto-3D) to a 150M-point client cloud processed on WWT's in-house HPC.

Alongside that, I designed and proposed REWIND, an onboarding and information-consolidation Claude skill that surfaces internal documentation and past project context for new hires, projected to cut ramp-up time by two weeks per hire. I also built and presented CARDINAL, an intern platform with features like coffee chat logs, calendars, and more, in addition to a manager-facing view to flag interns who may need support. CARDINAL placed 2nd out of 19 intern teams, presented to 100+ people.

Agentic AI Intern — Remote

LLM evaluation · Call classification · Six Sigma

I validated AI-generated call transcript summaries for accuracy and data privacy compliance, and developed call classification labels to improve model categorization across two production LLM systems (Quality AI and Agent-Assist). My contributions fed into an agentic AI initiative projected to reduce customer service call times by 15 seconds across their high-volume call centers.

I also pitched a customer engagement forum concept to Mr. Cooper C-Suite using Six Sigma techniques, receiving positive feedback on its potential as a retention and satisfaction tool.

OptMAI Lab — Texas A&M University

PyTorch · 3D ResNet · Partial AUC · HPRC GPU clusters

My Honors thesis focuses on predicting whether a patient with Mild Cognitive Impairment (MCI) will convert to Alzheimer's Disease, using 3D brain MRI from the ADNI dataset. The core research question is whether partial AUC, the metric clinicians actually optimize for when triaging high-risk patients, can be maximized directly as a training loss, rather than using cross-entropy as a proxy and hoping AUC follows.

The motivation is clinical: a model trained to minimize cross-entropy may not rank the highest-risk patients at the top of the curve where it matters most. By treating partial AUC maximization as the training objective directly, the goal is to align what the model learns with what a doctor would actually use it for.

OptMAI Lab — Texas A&M University

PyTorch · Contrastive learning · 3D ResNet · SLURM / HPRC

I conducted hybrid-loss experiments comparing cross-entropy against supervised contrastive loss for classifying 3D brain MRI scans into Alzheimer's Disease, MCI, and Normal Cognition using the ADNI dataset. The key finding: SupCon lifted recall on MCI (the hardest class, and clinically the most important to catch) by up to 23.5 percentage points, and AUC by 3.5 pp.

I built preprocessing pipelines for the volumetric MRI data and associated tabular data, trained on TAMU's HPRC GPU clusters via SLURM, and initialized the 3D ResNet backbone from MedicalNet-pretrained weights using transfer learning.

Data Engineering Research Assistant — Texas A&M University

Python · LiDAR · Optical imagery · Multi-modal deep learning

I developed data pipelines to fuse optical imagery and LiDAR point cloud data across 90+ field samples for multi-modal deep learning-based biomass estimation and mapping. The two modalities are complementary, optical captures surface color and texture while LiDAR provides structural depth, and fusing them produced richer representations for estimating crop biomass than either alone.

Aggie Data Science Club — Jan 2025 – Apr 2025

Python · Pandas · Snowflake · Scikit-learn · SciPy

Created a forecasting pipeline from millions of American Airlines bag-scan records, predicting the bag-load curve roughly two weeks out from flight-schedule features. The goal was to give ground operations enough lead time for proactive staffing rather than reactive scrambling on heavy travel days.

Signed NDA

Software Engineering Course — Aug 2025 – Dec 2025

Ruby on Rails · PostgreSQL · Docker · GitHub Actions · Agile/Scrum

Led a 4-person team as Product Owner to ship a full-stack Rails platform for a real client: the Vertical Flight Society at Texas A&M. Features included application workflows, dynamic sponsor displays, tagging, filtering, and user authentication, delivered across 3 sprints with 30+ standups. Deployed with Docker and GitHub Actions CI/CD.

Hackathon

Python · TensorFlow · Keras · Computer vision

A three-layer CNN built with Python, Keras, and TensorFlow to detect driver drowsiness from facial images, achieving over 80% accuracy on a curated dataset of 10,000 images. The business model was recognized as a top-10 hackathon project focused on vehicle safety.

Hackathon

React · OpenAI API · Generative AI · FinTech

A React web application that helps users manage and optimize their subscription expenses, using the OpenAI GPT API to give financial advice based on user data. Features include local data handling, AI-powered subscription analysis, and an interface for viewing spending insights.

Personal project

JavaScript · Spotify API · Render · Data visualization

A web application using the Spotify API to visualize user listening patterns (top songs, artists, genres, and temporal trends) with comparative analytics to benchmark personal preferences against global trends.

Software Engineering Course

TypeScript · React · PostgreSQL · Accessibility

A point-of-sale system for boba shops with full accessibility support across cashier, manager, and customer interfaces. Implemented business analytics including X and Z reports for tracking sales and inventory.