I build AI-enabled tools and check that they work.

The problem

My econometrics course used this data for one problem set: a few regressions of loan approval on race and finances. I wanted to know how far the answer moves with what you control for, and whether a model that never sees race ends up treating applicants equally.

What I built

• Linear probability, probit and logit models with marginal effects, a ladder of controls, race interactions and an Oaxaca-Blinder decomposition of the gap.
• Approval models trained without race, measured for demographic parity, equalized odds and calibration by group, plus how much race the other inputs carry.
• Mitigation trade-offs and a label-bias what-if, since the "approved" label is the decision being audited.

What it found

White applicants were approved 90.8% of the time and Black or Hispanic applicants 70.8%. About 63% of that gap isn't accounted for by the recorded finances and credit histories, which fits discrimination but also things the files don't record. A model trained without race kept a 21-point gap.

How I tested it

The course's models reproduce the instructor's Stata output to every printed digit (81 of 81 numbers), the fairness metrics are tested against hand-worked cases, and every push rebuilds the results and fails if they drift.

The problem

My mom noticed three data centers going up within a mile of my little brother's middle school in about six months, and our family had no idea any of them were coming. The existing maps rely largely on community reports, so they aren't pulling from the best data.

What I built

• A pipeline that pulls DEQ's daily air-permit records (every data center needs one for its backup generators), names each site's county or city, and adds permit dates.
• Search by ZIP code, school (all 2,159 Virginia public schools, matching the names people actually use) or a pin, within 1 to 30 miles. It all runs in the browser, so nothing a visitor types leaves their device.
• A daily automated refresh that refuses to publish if a source comes back broken.

What it found

Four data centers are within a mile of my brother's school, two of them still planned. Across Virginia, DEQ issued 44 data center air permits in 2026 through September, up from 34 in all of 2025.

How I tested it

Tests check that the browser's distance math matches the Python pipeline to a millionth of a mile, that "Rachel Carson" finds the school the federal data calls "Carson Middle", and that the build stops on bad data. Where two official sources disagree on a site's locality, the map shows both.

The problem

Ten courses of material sat in scattered formats, so finding how one idea connected across classes meant re-reading everything.

What I built

• Converters for PDFs, slide decks, spreadsheets and video transcripts into markdown with metadata (700+ documents across 10 courses).
• An image workflow where a local vision-language model classifies each page as a graph, diagram, table or text and describes the figures.
• Automated experiment loops that benchmark model, prompt and image-size variants against a hand-checked set and adopt a change only when it wins.
• A keyword search prototype over 7,000+ text chunks, the foundation for question answering across courses.

How I tested it

I labelled a 19-page set by hand and scored every variant against it. Page-type accuracy rose from 71% to 100% once the prompt asked for the page type on its own line. The table shows the leaderboard from one report.

Bars are drawn from 0.80 to 1.00 so the differences are visible. Prompt v5 looked reasonable but lowered the score because it broke slides that mix figures and text, so I did not adopt it.

What I learned

Several apparent model failures were really bugs in my own scoring code. I now read the raw outputs before I trust a metric.

The problem

UVA publishes its capital plan every year, but each year's tables use a different layout and projects change names between years, so following one project's budget over time is hard.

What I built

• A parser that turns six years of plan PDFs into one table of 286 project rows, each reconciling to the dollar.
• Linking that follows each project through renames: exact and fuzzy name matching plus hand-checked overrides.
• The cash-flow model from my real estate course (NPV, IRR), with sliders for overruns, delays and growth rates.

What I found

State funding fell from 26% of the plan to about 10% while debt rose to a 58% peak. Of 64 projects with more than one budget, 23 grew, by a median of 17%.

How I tested it

The model reproduces my course's worked example to the dollar, the browser version matches the Python version on 26 cases, and every push retests the site before it deploys.

The problem

I was managing more than 50 applications, each with its own deadline, requirements and follow-up date. Notes in scattered documents were not going to hold.

What I built

• A Python API and a single-page front end that saves every change to a JSON file I can review in git.
• Forms that reject invalid dates, scores and statuses.
• Views for what needs attention now: submissions to confirm, follow-ups due and upcoming deadlines.
• Controls against cross-site and DNS-rebinding requests: loopback-only binding, Host checks and JSON-only writes.

How I tested it

Automated unit tests cover validation, saving and the request checks. I built it with AI coding assistants and reviewed their work through the tests, which caught a date-validation bug that would have dropped connections.

The problem

Announcements were missed and there was no single, reliable record of members, because everything lived in Flare, Snapchat and GroupMe.

What I did

I gathered requirements from members, built the platform and released it in small steps. Members raised concerns about cost, time to adopt and design, and I addressed them by prioritizing their feedback.

The result

Other brothers now contribute to the project, it supports alumni connections, and the membership database is being extended to help manage house access.