I'm a Data Scientist with an M.S. in Information Science from UT Austin, combining business thinking with data science to solve real-world problems. My experience spans data analytics, machine learning, deep learning, and AI/LLM applications, with a focus on building practical, data-driven solutions.

📍 Austin, TX  ·  🔗 LinkedIn

MS capstone, UT Austin (2026)

Analyzed the full EMS call lifecycle in NYC using FDNY dispatch data (~28.7M records). Engineered temporal, cyclical, spatial, and rolling-window features, compared 8 models (linear regression, random forest, gradient boosting, XGBoost, LightGBM, MLP, LSTM, and an ensemble), and used a DeepSurv survival model to estimate the probability of delayed ambulance arrival in different scenarios, such as weekday rush hour vs. weekend night.

Findings: Tree-based ensembles outperformed linear and LSTM models. Hospital turnaround (median ≈ 40 min) and on-scene time (≈ 22 min) drive total cycle time far more than dispatch or travel.

Python Machine Learning Deep Learning Random Forest LSTM MLP DeepSurv pandas scikit-learn XGBoost LightGBM TensorFlow Survival analysis

Data Wrangling, UT Austin (2025) Tested whether college football programs that spend more win more, using three datasets (a 710-column federal financial file, bowl results, and season results), joined across inconsistent school names with an AI-assisted, manually verified lookup table. The same pipeline is implemented in R, SQL (DuckDB), and Python to compare the approaches.

Findings: Higher-spending programs have a higher median win percentage and more consistent results, and the gap between the top-25 and bottom-25 spenders holds across seasons from 2004 to 2021.

R SQL DuckDB Python Quarto

Team project, Deep Learning and Multimodal Systems

Detects AI-generated speech using Wav2Vec2 embeddings and Mel-spectrogram features on 2-second audio clips. Compared a baseline MLP, a dropout-regularized MLP, and a spectrogram-based model. Test accuracy on the Fake-or-Real dataset improved from about 53% for the first baseline to about 92% for the best model.

Python Machine Learning Deep Learning MLP Wav2Vec2 PyTorch Hugging Face Transformers Librosa

Case study prototype

An end-to-end pipeline that automates a four-stage invoice workflow (ingest → validate → approve → pay), reading five file formats. An LLM handles messy text extraction and approval reasoning with a self-critique pass, while the costly business rules (such as the $10K scrutiny threshold) are enforced in plain Python. A batch runner produces an audit-friendly summary of every decision.

Python LLM APIs SQLite AI-Agents sqlite

• Languages: Python, SQL, R, Excel
• Analysis and modeling: regression, tree-based ensembles, deep learning, survival analysis, feature engineering
• Libraries and tools: pandas, scikit-learn,NumPy, XGBoost, LightGBM, TensorFlow, PyTorch, tidyverse, DuckDB, SQLite, Quarto
• Engineering & Tools: Git · GitHub · APIs · Jupyter · Docker
• Visualization: Tableau · Matplotlib · Plotly

Exploring opportunities in data science, analytics, machine learning, AI, and data-driven product/engineering roles.