Every industrial roof in a country, assessed without a site visit

A solar business knows exactly which companies are worth calling. It just cannot find out which ones they are.

The qualifying question is simple to state: does this company have a large, empty, usable roof — and does it use enough energy to justify covering it — and has somebody already done it? Three facts. Any salesperson can name them.

Getting those three facts about one company takes a morning. Someone opens a satellite view and squints at the roof. Someone else looks up the financials and guesses at consumption from the sector. Somebody checks whether panels are already there. Multiply that by every industrial company in a country and the arithmetic collapses.

So it does not get done. The market goes unassessed, and the sales team works from the companies it happens to know — which is a tiny and arbitrary slice of the ones worth calling. The constraint is not knowing what to ask. It is that the answer costs a morning per company.

An assessment of every industrial company in Slovenia — roof, energy profile and subsidy history — built entirely from desk data, and delivered as a list the sales team can filter and call.

Roughly 30,000 companies. No site visits.

For each one, the system locates the company's buildings, pulls 25 cm orthophoto imagery from the Surveying and Mapping Authority of the Republic of Slovenia (GURS), and reads the roof: usable area, whether panels are already installed, and how much capacity could realistically be fitted. It reads the company's website to establish what they actually make. It builds an energy-consumption profile from sector intensity scaled to company size, refined against what the business describes about itself. It joins in Slovenia's public renewable-energy support register and the de-minimis state-aid register, so a company that has already taken subsidy is visible before anyone calls it.

The result is a single profile per company across 79 fields, an annotated image of the roof, and an application where a representative sets their own filters — roof area above this, no existing panels, energy tier above that — and gets a ranked call list.

The decision that makes the numbers trustworthy

The roof measured is the company's own building — not the industrial zone it sits in.

This is the part that determines whether the whole dataset is usable. Industrial companies cluster in shared zones, and roof detection run naively returns the entire plot: every tenant inherits the whole cluster's roof.

The difference is not marginal. One tenant — a company of around €2M revenue — would have been presented as 92,513 m² of roof and 7,794 kWp of potential. Its own building is 3,289 m² and 86 kWp. A representative who called that company quoting the first number would have lost the meeting in the first minute.

So building attribution is handled geometrically rather than by a model: each roof area is partitioned to its nearest company and classified by scope. It has to be exact and explainable, and this client is rightly intolerant of fabricated precision. That is also why the energy baseline is a fixed coefficient table rather than a guess — a model refines it, but never invents the floor.

Because it lost. Before any of this was built, the roof-reading step was decided by testing the alternatives against each other on real Slovenian industrial roofs.

Google's Solar API — the purpose-built tool for exactly this job — was tested alongside Gemini, GPT and several open vision-language models. All of them were beaten by Claude Sonnet 5, which is what the system uses.

What that comparison is, and is not. It was judged by a person looking at each model's output against the imagery and deciding which reading was right. It is a selection method for one use case — large, irregular, partially occupied industrial roofs in one country — not a scored benchmark, and not a general claim about any of these products. On residential rooftops in a well-mapped city, the answer may well be different.

We publish it because the question is the obvious one to ask, and because the answer was not what we expected going in. The purpose-built tool is usually the right default. Here it was not, and finding that out took a structured comparison rather than an assumption.

Delivered 28 August 2026:

And the timeline is the part that is hard to believe. The work was won on 20 August. The application was live on 22 August. The complete data package — ~30,000 companies and 93 GB of imagery — was delivered on 28 August.

Eight days from winning the work to a sales team with a ranked national call list. That was possible only because the classification base, the enrichment pipeline and the roof detection already existed. What this project added was the honest building-level attribution, the energy profiling and the delivery.