Gramm did not start in energy. We started in semiconductor manufacturing. Anomaly detection, root cause analysis, process deviation scoring. The kind of work where a missed defect costs six figures.
In early 2025, we came across the CEC's grant program for grid forecasting. The brief described a familiar problem. Predict a noisy time series with weather-driven nonlinearities. Quantify your uncertainty. Do it fast enough for operators to act.
That is a manufacturing quality problem wearing a different hat.
In a fab, you watch process variables for the moment something drifts. In a grid, you watch load and weather for the moment demand departs from its expected shape. The math is the same. A missed defect scraps a wafer. A missed load spike triggers emergency dispatch.
We studied the existing landscape. The models in production were not bad. They were just old. Tuned for average conditions. Fine on a Tuesday in April. Broken during a heat wave in August.
The normal case is easy. The tail is where it matters. Our tools were built for tails.
The first model ran on California data and beat the CAISO baseline. We ran Texas. Then PJM. Then MISO. The pattern held everywhere.
Next: Why we invented a new accuracy metric before building the model