(01) — AI in my workflow
Where it genuinely helps, and where it is just noise
AI has not changed what I decide; it has changed how long it takes me to get to the decision. These are the five places it enters my process, and in every one of them the final review is mine.
- Research synthesis Twelve interviews, four hundred support tickets and an analytics spreadsheet become a list of patterns in an afternoon rather than a week. Then I go back to the sources and check each pattern: what I cannot find in the original material gets dropped.
- Interface variant exploration Six ways of solving the same settings table, in the time one used to take. They exist to discard quickly, which is what variants are for; the one that survives I draw myself.
- Code scaffolding and refactors The skeleton of an Angular or React component, the conversion of a layout into components, the repetitive migration of a thousand files. I read all of it before it enters the repository.
- A first accessibility pass It finds the missing
aria-labeland the short contrast in seconds. It does not replace testing with a screen reader, which is the part that actually decides. - Interface copy The error messages, empty states and labels nobody wants to write. Ten versions to pick a tone, then a rewrite by hand so it sounds like the brand and not like a template.
What AI does not do in my process
It does not decide the information architecture, because that needs knowing what the content means to the business publishing it. It does not sign research: if a fact is not in the source, it does not go in the report. It does not ship code I have not read line by line. And it does not make the brand identity, which still comes from understanding one specific client rather than a statistical average of all the others.
This is a working stance, not a decoration: it comes from thirteen years signing ceramic reports for a CSIC archaeological mission, where a fact has to be checkable and, if something is not known, you say so.
(02) — Designing for AI
An AI interface is a different design problem
A normal product answers the same thing twice. An AI one does not. That breaks the three things almost all interface design leans on: that the result is predictable, that the error is clear, and that the user knows whose fault it is when something goes wrong.
Designing it well is mostly designing the edges: how you show that an answer is an estimate and not a fact, how it gets corrected without starting over, what happens while it generates, what happens when generation fails halfway, and how you say — without small print — what is being done with the data of the person asking.
And then there is the brand. Almost every AI company arrives with the same problem: they want to look advanced without looking interchangeable, in a sector where everybody uses the same purple gradient.
(03) — The case
KAI · Kallisto AI
Identity for an artificial intelligence company working on drones, with three product lines under one brand.
See the caseSee the case ↗The brief was not a logo: it was a system. The parent brand had to work on its own and take three surnames — SKY, SHIELD and SEG — without turning into three separate brands.
Two routes were presented: a circuit trace drawn out of the word itself, and a neural network symbol that can sit above or beside it. Each was tested in positive and negative, horizontal and vertical, and with all three sub-brands attached, which is where systems break. The second one won.
(FAQ) — Frequently asked
What people ask before hiring me
Does this fit what you need?
Tell me what you are building and I will tell you plainly whether I am the right person for it. I reply within one working day, Madrid time.
Let's talkLet's talk ↗