Selected work in three areas: equipment designs running in a power plant, software I built for school and business, and machine learning models for the ERCOT grid. Each entry has a short write-up and screenshots.
Invented a new style of flame scanner to overcome the poor line of sight that plagues secondary flame scanners on gas turbines with DLN1 combustion systems. The newly designed units are currently in use on running turbines and are being manufactured by BFI Automation. I presented the design at the NAES O&M Conference in March 2026.
Designed an external lube oil cooling unit to keep oil and bearing temperatures within specification during hot weather runs. Applied heat transfer and fluid dynamics to calculate the required heat removal in BTU/hr, then worked within a hard constraint: no new penetrations into the lube oil tank, so flow was limited to the tank's existing openings. Calculated the maximum flow the piping could carry without cavitation and used it to size the pump, then the fin fan and heat exchanger. An engineering company produced drawings from the design, and the unit was installed on six GE Frame 7B units. On a 90°F day, lube oil temperatures dropped 20 to 25°F, varying slightly between units.
While working through corporate finance and managerial accounting in my MBA, I kept rebuilding the same formulas in Excel for each problem set. In an effort to stop repeating that work, I built a desktop app that holds about 230 corporate finance formulas in one place: time value of money, NPV and IRR, capital budgeting, cost of capital, DuPont, EVA, and free cash flow, along with the costing, CVP, and variance formulas from managerial accounting.
The feature that makes it useful is that every formula solves for any unknown. Enter the variables you have, leave one blank, and it solves for that one. It also produces full amortization schedules. I designed and specified the app and directed its AI-assisted development with Claude Code, then tested every formula against the textbook examples from Brealey, Myers, and Marcus. It builds for both Windows and Mac.
Python · Tkinter · Windows and Mac builds
This started as a price spike predictor proposal for an MBA course and grew into a full dashboard. It runs XGBoost models on a year of real ERCOT data pulled from the gridstatus and ERCOT APIs, EIA natural gas prices, and Open-Meteo weather.
The models answer two questions a peaking plant cares about. The first is whether real-time prices will spike above $500/MWh one to four hours ahead, which it predicts with an AUC of 0.76. The second is whether the grid will enter an emergency condition one to six hours ahead, which it predicts with an AUC of about 0.99 and recall between 0.79 and 0.94. I backtested the emergency model against Winter Storm Uri. The dashboard also tracks wind and solar forecast error and uses SHAP to show which inputs drove each prediction, so an operator can see why the model thinks what it thinks instead of taking a number on faith.
Python · XGBoost · scikit-learn · SHAP · Streamlit · ERCOT, EIA, and Open-Meteo APIs
I built a complete high school chemistry course as an interactive web app aligned to the Texas standards, TEKS §112.43. Each unit has a lesson, worked examples, and auto-graded practice problems, so a student gets immediate feedback instead of waiting on a graded worksheet.
The AI work is in the content pipeline. I specified the standards coverage and the problem formats, then used Claude to generate the lessons and problem banks unit by unit and checked each one against the standard before it went in. This is the same approach I use on the plant simulator at work: define the structure and the acceptance criteria, let the model fill it in, and verify the output.
HTML/CSS/JavaScript · Claude-generated content · Static hosting
