Emily Stepro

Schedule Master is useful for seeing who is booked and when. What is harder to see is the bigger picture of instructor capacity across the club.

I am building a dashboard to help with that. It looks at each CFI’s activity, teaching areas, specialty qualifications, and overall schedule load. That includes private pilot training, instrument training, Cirrus, complex, tailwheel, flight reviews, IPCs, and other proficiency work.

The purpose is to give the Chief CFI and club operations a clearer way to understand where instructor coverage is strong, where it is thin, and where member demand may be outpacing availability.

Built with Python and n8n, using data from the Schedule Master API. The information already exists in the system. This project is about organizing it in a way that helps the club make better decisions.

Students usually have a CFI paying attention to their progress. If they stop flying for a while, cancel several lessons, or begin to lose momentum, someone is likely to notice.

Certificated members are different. They may fly regularly for a season, then less often as work, family, weather, or life gets in the way. The gaps can build gradually, and there is not always a clear signal that it may be time for a reset.

I am building a monitor that looks at certificated pilots through two lenses. One is currency, where FAA or club requirements may be approaching a lapse. The other is proficiency, where recent flying patterns may be drifting from a pilot’s own baseline, including gaps in instrument, night, or varied flying conditions.

The goal is to make the next step easier, not to create noise. When something looks worth attention, the system can surface a relevant Blindspot article, a WINGS activity, or a simple path to schedule a proficiency flight.

Pilots get their own view. The safety team gets a roster view. Built with Python and n8n.

I kept running into the same problem with prompts: I’d write one, it seemed to work, I’d move on. But I never really knew if it was actually better than the version I didn’t try.

So I built an A/B testing system, using aviation weather as the testing ground. It pulls live METAR and forecast data, runs two prompt versions against the same real conditions, and scores each response with an LLM judge on accuracy, clarity, actionability, and tone. Results build up across multiple trials so you’re comparing patterns, not a single output.

I wanted something I’d actually trust for picking prompts. A repeatable process I could point to and say, this one’s better, and here’s why.