TH Arquitectos: Certification Documentation Engine

2026 · 3 months · TH Arquitectos

AI Design Technologist, Computational Designer, Full-Stack Developer, Technical Lead

Local-first desktop app that takes an Arquímedes construction budget and a folder of technical sheets, and gives back the official sustainability certification template pre-filled, work that used to eat days on every certified project. First of several automations I'm building with TH Arquitectos.

The problem

Sustainability certification at an architecture studio means cross-referencing an Arquímedes construction budget against a folder of supplier datasheets, line by line, by hand, on every certified project. Days of work each time.

The constraint

Real budgets carry prices, supplier names and project metadata. That data cannot leave the studio, which rules out a cloud service by default. And the two sides speak different languages: supplier trade names on one side, the certification body's controlled vocabulary on the other.

What I built

A local-first desktop application running a five-step pipeline: parse, match against a curated master table, AI on the technical datasheets, AI on the long tail, write the Excel template. Streamed over SSE, every step independent and testable on its own. A table of around 24 known materials resolves most line items with no LLM in the loop, so the AI only sees what falls through, which keeps cost, latency and confident wrong answers in check. Prompts are served by a Vercel Edge Function and cached locally with a bundled JSON fallback, so changing a prompt does not mean rebuilding the binary.

Role

Independent consultant. Discovery with the studio, architecture, the full build, packaging and delivery.

Outcome

In use at the studio. A certification that meant days of manual cross-referencing per project now runs in a single pass, and every file stays on the studio's own machines, which is what made it viable for client documentation at all.

Lessons learned

• Local-first as a design constraint: real Arquímedes budgets carry prices, supplier names and project metadata. That doesn't belong on someone else's API by default. One rule, and the whole architecture falls into place: cloud is opt-in, never opt-out.
• Curated database first, AI for the long tail: a small table of around 24 known materials handles most line items instantly, with no LLM in the loop. The AI only sees what falls through, which keeps cost, latency and confident wrong answers in check.
• The 5-step pipeline as a reusable skeleton: parse, match against the master table, AI on technical sheets, AI on the long tail, write Excel. Streamed over SSE, every step independent and testable on its own.

Final reflection

This is the first thing I've built with TH Arquitectos, and it won't be the last. It picks up the idea I kept coming back to in Stairs and Timber: bake the expert knowledge into the tool itself, so the rules live in the artefact and not in someone's head.

The interesting bit wasn't the maths; it was the semantic gap between two ways of describing the same building: the supplier-driven Spanish that ends up in an Arquímedes budget on one side, the controlled vocabulary the certification body expects on the other. And certification carries professional liability: automation that removes control is a harder sell than automation that speeds it up.