Quantaris

The problem

Producing a construction budget from a floor plan is skilled, slow work: read the drawing, identify the rooms, measure the areas, count the materials, price them. It takes days per project, and an error only surfaces after the materials have been bought.

The product started as a prototype that proved the idea worked. Turning it into something a construction firm could run its money through was a different engineering problem entirely.

Rebuilding the foundation

First, infrastructure the business owned: its own PostgreSQL in São Paulo, 47 tables under row-level security, deployment on Cloudflare Workers behind its own domain, and a deploy that is documented and reproducible rather than clicked. Continuous integration runs typecheck, lint, tests and build on every pull request.

The AI moved to a direct API integration with streaming, off the prototype's shared gateway. Query volume per screen went from around 114 to under 10.

The bug that mattered most

In the prototype, running a plan analysis could destroy work the user had entered by hand. That is the kind of defect that ends a product: it does not corrupt data visibly, it deletes effort, and the user finds out later.

It was rebuilt as staging, diff and transactional apply — the analysis proposes, the difference is shown, and the change lands atomically or not at all. An adversarial test proves isolation between projects, so one analysis can never reach into another.

Multi-tenant, done properly

Authorisation was rebuilt around organisations rather than individual owners: an organisation identifier across 29 tables, role functions, and 56 row-level security policies migrated with a transitional trigger so production never broke mid-migration — both branches of the fallback validated before the old path was removed.

Making the AI economics knowable

Cost per analysis was measured rather than estimated: R$ 0.13–0.26, well under the original projection. Getting there surfaced two things worth knowing — that an empty model response was passing validation and being billed as a successful analysis, now failing loudly and charging nothing; and that Gemini's cost has to be computed from total tokens minus input tokens, because reasoning tokens are billed but invisible if you only read the output count.