What we've builtEnergy certificates
A dossier read in ten minutes, every page
Certificate dossiers took a reviewer thirty to eighty minutes. We built an engine that reads all of it, runs about 140 checks and shows its evidence.
The setup
A certificate dossier is a stack of PDFs: quotes, invoices, attestations, site photos, company registrations, technical sheets. Conformity means every value agrees with every other, the photos show what they claim, the installer is who they say they are. Reviewers did it by eye. Verdicts varied by reviewer and by the hour of the day.
What we did
We split reading from deciding. A language model classifies each document and extracts its fields into structured data. It does not decide anything. Deciding is done by explicit rules, about a hundred and forty of them: company identity against the public registry, names and addresses consistent across documents, quantities against technical sheets, photo timestamps and locations, model references against the approved list. A cross-document comparison catches what no single-document check can.
Every check carries its evidence with a link to the page. A reviewer verifies in seconds instead of re-reading. Findings are weighted, critical to minor, into one confidence score.
The engine got good the way a reviewer does: by being wrong on real dossiers and corrected. Sixty-odd iterations on a set of real cases took it from prompt tweaks to a rules architecture with snapshot tests, so a fix for one dossier cannot silently break another. Each correction became a rule or a test. The same architecture now checks grant dossiers too.
What changed
Ten minutes, not an hour
Read, checked, evidence attached.
Same rules, every dossier
And every rule has a test.
Auditable by design
The model proposes, the rules decide, the page is one click away.