Computer vision for electrical panels

Goal

Turn a promising computer vision model into a product feature people could actually rely on — one that reads a photo of a home’s electrical panel and computes its available load.

The problem

Electrical load calculations gate every high-power installation: EV chargers, heat pumps, the rest of home electrification. Done by hand, they’re slow and inconsistent. The model existed in research form, but a research model and a production feature are different things. It needed a trustworthy training set, a measured accuracy bar, and a workflow that held up in the field.

What I did

I worked with a central AI team and a small product engineering team to take the model to production.

• Built the ground-truth dataset. I ran the subject-matter-expert labeling workflow — partnering with licensed electricians to label hundreds of panel photos into the training set the model learned from.
• Set the accuracy bar. Before launch, I ran precision/recall evaluation to measure how the model actually performed, so the decision to ship was based on evidence, not optimism.
• Shipped it as a feature. The model now powers a load-calculator used as a standard step in installation work.

The result: load calculations that used to take significant manual effort run in a fraction of the time, with a measured accuracy bar behind them.