2026
Python, LLM agents, USCIS disclosure data
In progress
OPT Career Intelligence Agent
International students apply to employers who have never sponsored a visa, because nobody tells them which ones do. The information is public, in federal H-1B disclosure filings, but it is not in a form anyone can use when they are applying to forty companies in a week.
This is an agentic system that reads those filings and answers the question directly. It is my capstone, and I am opening it to the public this year.
CapstoneMaster of science in data science and analytics, 2026
2026
PyTorch
A transformer language model, built from scratch
Eight weeks as an Open Avenues Build Fellow, building and training a GPT-style model end to end — the tokenizer, byte-pair encoding, the training loop, and deployment. I wanted to know how the thing works at the level of the code rather than the level of the diagram.
End to endTokenizer through deployment, nothing borrowed
2026
R, tidymodels
Family engagement and equity
Generalized linear models across 25,391 students, asking whether family engagement protects students differently depending on income.
It does. The protective effect is 18 percent stronger for low-income students, at p = .009, and it held across four separate model specifications, which is the part that made me believe it.
25,391students in the sample
2025
R, Quarto, Census ACS
Poverty and gun violence
Census data merged with more than 100,000 incidents across all fifty states, tested with permutation tests, a 10,000-sample bootstrap, and Welch's t-test.
Poverty predicts the rate of incidents. Population does not, which is the assumption most people start with.
50 statesand more than 100,000 incidents
2025
Python, scikit-learn, mlxtend
Alzheimer's pattern discovery
A knowledge-discovery pipeline over 2,149 patients, reaching 93.8 percent accuracy with an interpretable decision tree rather than a model nobody can read, and validated three different ways.
In clinical work the interpretability is not a nice extra. A doctor has to be able to see why.
93.8%accuracy, from a tree a clinician can read
2025
Python, SQL, BigQuery, Looker
Weather ETL pipeline
A five-phase pipeline moving more than 96,000 records into SQLite and BigQuery, with a live Looker dashboard on the end of it. Ordinary engineering, built properly, which is most of what the job really is.
96,000records, moved and monitored