Finding the megawatts that are already on the grid
The problem
The stranded-power thesis only works if you can name specific addresses. Public data will not tell you which buildings are underused, so the work is inference across hundreds of thousands of footprints, and the failure mode is a language model inventing a plausible owner and phone number.
What we built
A region-pluggable ETL pulls building footprints, substations, transmission lines, interconnection queues, and layoff notices. ERCOT came first and PJM second. Adding a grid operator is configuration, not a rewrite.
A PostGIS matcher estimates capacity from footprint area, finds the nearest substation by nearest-neighbor search, assigns an archetype, and ranks the result. Enrichment costs money, so only sites above a score threshold ever reach a model.
The enrichment agent researches owners and contacts with live web search, and the server drops any field that does not arrive carrying a real source URL. A site with no verifiable owner shows no owner. This runs server-side; it is not a request made in a prompt.
A monthly ledger caps enrichment spend with a protected per-region floor, so one region draining the global pool cannot starve another. The budget is a database table, which means it survives a restart and can be audited.
Server-rendered dossiers with aerial imagery, grid context, sourced facts, and structured data give each candidate a shareable, indexable URL. The map is for exploring; the pages are for selling.
What this demonstrates
Retrieval and enrichment pipelines with fabrication guards, spend ceilings, and provenance on every claim.
Get in touch
Tell us what you are trying to build and what is in the way. We will tell you early if we are not the right fit.