The problem
Visibility was useful. It was not enough.
The first version did exactly what a good reporting system should: it brought solar generation, battery state, household demand, half-hourly energy prices and financial projections into one live view.
It alerted when prices turned negative. It tracked import, export and the year-end position. It made the energy system legible.
But the action still sat elsewhere. The cloud API used for control was known to be unreliable. A successful command could be followed by failed reads, leaving the system unable to confirm state or safely revert. And tomorrow’s schedule still depended on assumptions about when the house used energy.
The next step was not “more AI”. It was a trustworthy loop: read locally, act locally, verify the result, fail safe—and use real history to decide what should happen next.
