Darrell Burch

A public website that grades every member of Congress on attendance and legislative record — and updates, deploys and tests itself every night with no manual steps.

• Built a nightly GitHub Actions pipeline that pulls data from three public sources, rebuilds the site's data, publishes to GitHub Pages and then runs Playwright smoke tests against the live site.
• Made the build fail safe: requests retry with backoff, and if more than 5% of records fail to load the run stops and the site keeps the previous day's data.
• Converted 11 recorded Selenium IDE tests to Playwright for parallel cross-browser runs, with a trace captured on retry and an HTML report saved from every run.
• Wrote tests that check the pipeline's output as well as the page: record counts, duplicate IDs, internal consistency and data freshness.
• Added a JMeter plan that reads the member pages from the sitemap, runs a pass of smoke checks, then simulates visitors browsing the live site.

A test suite for an AI feature: a tool that answers plain-English questions about members of Congress from the Report Card's data, and the evaluation that measures whether its answers can be trusted.

• Designed the cases around how an AI feature fails: inventing figures, answering about people who are not in the data, giving political opinions, and obeying instructions hidden in a question.
• Graded facts with code and judgement calls with a second AI model, and tested that judge on answers with known verdicts before trusting it.
• Computed every expected answer from the live data at run time, so the suite stays correct as the data changes each night.
• Kept failures of the plumbing apart from failures of the model: an API error or timeout is logged and retried, never scored as a wrong answer.
• Proved the graders reject wrong answers that look right, and added a GitHub Actions workflow that runs the free checks on every change.

A reusable starting point for UI and API test automation, built the way I would set one up for a new team.

• Page objects hold locators and actions only; assertions stay in the specs, so a UI change is fixed in one place.
• Custom fixtures inject page objects and seed test data through the API, and a setup project signs in once and reuses the saved session.
• UI and API specs run in the same suite, with smoke tests tagged so they can run alone.
• A GitHub Actions workflow runs the suite on every push and pull request and keeps the HTML report; failures keep a trace, screenshot and video.

A Flask app that builds FanDuel lineups with a linear-programming optimizer and predicts game outcomes, covered by a pytest suite at every layer.

• 66 pytest tests cover the optimizer's rules, the data import, the matchup filters, the Vegas-odds strategy and the game predictor.
• Browser tests use pytest-playwright against the real app, started in a background thread, instead of a mocked client.
• Fixtures isolate each test's uploaded data, so a test run never changes the app's real data files.