01 — Micblau PiscinesBrand, site and technical assistant
The knowledge was in 67 documents
Water treatment for public and private pools. Brand, a site in three languages, and a technical assistant that answers from the company's own documentation and always cites where it came from.
The facts
- Client
- Micblau Piscines, Vilassar de Dalt, Spain.
- What we did
- Identity, trilingual site, transactional email, security, and a technical assistant with retrieval over their own documents.
- Status
- In production.
- Year
- 2026.
Water treatment is a trade made of technical sheets. How much product per cubic metre, what to look at when the water clouds over, what to do when somebody pours in the wrong thing: all of it is written down, and written well. The problem is that it is written across sixty-seven documents, and the person with the question picks up the phone.
The brief was not «add AI». It was to stop that knowledge — which already existed and was good — from depending on one particular person answering the phone.
What we built
Identity and a site in three languages. Catalan, Spanish and English written from the first line, not translated afterwards. Sixty-six routes, and eighty-six redirects from the old site so nothing that already ranked was lost.
The technical library, indexed. Sixty-seven documents split into one thousand one hundred and forty-eight chunks, with a relevance threshold calibrated against real questions. Every answer cites the sheet and the year it came from — and if it cannot find the source, it says so.
A photograph of the water as a question. The visible state of a pool is data the technician already uses. Here it is another input, not a demo trick.
Security and email, audited. A hard content policy, headers, rate limiting and transactional email on a verified domain. Audited in July 2026.
The numbers
- 67documents in the corpus
- 1,148indexed chunks
- 10/11hits on the control questions
- 89%prompt cache reuse
- $0.0018per query
- 66routes across three languages
Measured at the project's last verification, not estimated. If one changes, it changes here.
What nobody shows
The decision that took longest was not about the model, it was about the threshold. A system like this always finds something that resembles the question, and the temptation is to let it answer: it demos better. Here it was calibrated against real questions until the hits separated from the noise, and below that line the assistant says it does not know.
That is the part nobody shows, and it is half the work. In water treatment a confident wrong answer is not a product bug: a customer can end up putting the wrong thing in a pool. That is why every answer arrives with the sheet and the year beside it — so the last check stays with a person.
Next case
Let's talk
Tell us what you're building.