TRI NHAM

Data Science Research

Led the full EDA and modeling pipeline on the METABRIC breast cancer dataset (2,509 patients) with Dr. Pablo Guillén-Rondón at UHD — missing-data analysis, feature separation via Cohen's d & Phi coefficient, correlation heatmaps, and class balance review across 6 clinical targets. Presented at EGETC 2026, where it won the Best Paper Award.

Machine Learning Model Development

Engineered 18 clinical features and trained/compared Traditional ML, MLJAR AutoML, and TabNet under stratified 5-fold cross-validation with SHAP explainability. Gradient Boosting reached AUC 0.998 (NPI High Risk) and 0.952 (HER2+), with SHAP identifying lymph node positivity, NPI score, and tumor size as dominant predictors.

Exploratory Data Analysis (EDA)

Run systematic EDA across projects — missing-data audits, statistical feature separation (Cohen's d, Phi coefficient), correlation heatmaps, and class-balance checks — to guide feature engineering and modeling decisions.

Data Analysis

Delivered structured dataset analysis and reporting for client and internal stakeholders — from ETL-backed business insights at Cognixia to daily liquidity rate reporting at Capital One.

Data Pipeline Engineering

Built and maintained ETL pipelines across Databricks, Snowflake, and Spark at Capital One to support multi-source financial data analysis, reducing downstream errors by 25%.

Data Processing & Manipulation

Transformed and cleaned large tabular datasets — from Cognixia's client ETL workflows to feature engineering for the METABRIC research.

Backend & API Engineering

Translated business requirements into REST API-based data ingestion workflows at Capital One, connecting upstream systems to daily reporting pipelines.

Automation & Data Validation Testing

Built automated data validation pipelines and testing checks at Capital One, cutting manual rework by 40% and improving reporting confidence across teams.

End-to-End Project Delivery

Delivered an in-house replacement for a third-party liquidity reporting dependency at Capital One — owned end to end from design to production — saving the company millions in vendor costs.

Production Support & Bug Fixing

Rotated on-call for pipeline production support at Capital One. When pipelines failed, I dug through logs to isolate root cause — ECS timeout/bandwidth issues, schema mismatches, duplicate records — fixed what I could directly and escalated upstream file issues, keeping same-day ingestion on track for downstream teams.

AI/ML Development

Built DP Tuner, a differential privacy evaluation system that automates ε/δ accounting and hyperparameter tuning for synthetic health datasets.

Research Paper Contribution

First author of "Multi-Target Machine Learning for Breast Cancer Prognosis and Molecular Classification Using METABRIC", accepted and presented at EGETC 2026, where it won the Best Paper Award.

Software Engineering

Wrote and maintained production code spanning backend data systems at Capital One and a native mobile app (Histudy) — from data models to UI logic to automated tests.

iOS App Development

Built Histudy, a native iOS app that helps Vietnamese-speaking immigrants prepare for the U.S. citizenship civics test with a bilingual, AI-adaptive tutor and a spaced-repetition learning engine (SM-2), including unit tests for the learning engine and content integrity.

Bachelor of Computer Science

University of Houston

3.5

Master of Data Analytics

University of Houston-Downtown

4.0

Best Paper Award, EGETC 2026

Multi-Target Machine Learning for Breast Cancer Prognosis and Molecular Classification Using METABRIC, presented at the 5th International Conference on Electronic Governance with Emerging Technologies.

Race History

6 half marathons and 2 full marathons so far (Houston 2026, San Francisco 2026), plus 1 Ironman 70.3. Next up: Houston Marathon 2027. Full training log and every race is on Strava.

PhD Research Direction

Applying for Fall 2027 PhD programs with two research threads: bioinformatics, meaning ML on genomic and clinical data, and closed-loop neural interfaces, meaning interpretable, uncertainty-aware decoding of motor intention from EEG to time subthreshold stimulation, so an interface speeds up action while preserving the user's sense of agency.

Models only reach patients and users when they can be audited and trusted, so interpretability has to be designed in from the start. For noisy, subject-variable physiological signals, that also means calibration-light, personalized, privacy-preserving ML, validated with subject-grouped cross-validation and calibrated confidence.

Multi-target ML with SHAP explainability on METABRIC (Best Paper Award, EGETC 2026), NeuroTrack, an interpretable two-stage EEG motor-intention decoder, privacy-preserving synthetic data tuning with DP Tuner, and two-plus years of production data engineering.

ML for Breast Cancer Prognosis — METABRIC

Investigating whether machine learning can predict breast cancer prognosis and risk stratification from the METABRIC clinical and genomic dataset — 2,509 patients across 6 different clinical targets.

Combining engineered clinical features (NPI risk groups, LN flags, log-mutation, ER×HER2 interaction) with modern ML and AutoML approaches can match or exceed standard prognostic scoring, while SHAP explainability keeps the model clinically interpretable rather than a black box.

Led the full EDA pipeline (missing data analysis, Cohen's d and Phi coefficient feature separation, correlation heatmaps, class balance review), engineered 18 clinical features, ran stratified 5-fold CV comparisons across Traditional ML, MLJAR AutoML, and TabNet, and produced the SHAP explainability that identified lymph node positivity, NPI score, and tumor size as dominant predictors — Gradient Boosting reached AUC 0.998 (NPI High Risk) and 0.952 (HER2+). The paper was accepted and presented at EGETC 2026 (5th International Conference on Electronic Governance with Emerging Technologies, Springer CCIS), where it won the Best Paper Award.

DP Tuner

A research framework for automatically tuning the differential privacy budget (ε) in synthetic data generators — DP-TVAE and DP-CTGAN — designed for healthcare and HIPAA-aligned experiments.

Instead of picking an arbitrary privacy budget, you can search for the smallest ε that still preserves downstream model utility (e.g. AUROC ≥ target) — giving strong, provable privacy guarantees without over- or under-protecting patient data.

Built the framework end to end — the ε/δ tuning search loop, DP-TVAE / DP-CTGAN integration, and the utility-vs-privacy evaluation harness used to validate the results.

NeuroTrack — EEG Decoding

A two-stage, interpretable decoder for motor intention from non-invasive 64-channel EEG (86 subjects), built solo for the Rice Datathon 2026 NeuroTech track. Stage A separates real from imagined movement; stage B classifies the limb involved (left hand, right hand, both hands, both feet).

Physiologically grounded features and transparent models make EEG motor decoding interpretable: even when cross-subject accuracy is limited, you can still see which neural patterns drive each prediction and where on the scalp they live.

Built the whole pipeline solo: theta/mu/beta band power, hemispheric asymmetry, midline dominance and log-variance features for logistic-regression models, subject-grouped cross-validation to prevent leakage (~58% real vs. imagined against 50% chance; ~27–28% limb against 25% chance), and MNE-Python scalp topomaps to inspect the signal.

Stockd

An AI-powered stock thesis tracker — know when the story behind a stock changes before the price does.

Price alone doesn't tell you when your investment thesis has broken. Tracking the thesis itself — news, fundamentals, narrative shifts — surfaces the moment the story changes, ahead of the price catching up.

Built it end to end — product concept, backend, and the deployed web app.