AI agents that do real work
Almirant is the lab: orchestration, evals and cost per token. Almost everything I learn there ends up on the blog.
I'm F. Javier Pedrosa, PhD, and that line above is the scientific method applied to software and business. I'm also published academically as Francisco Javier Pedrosa Ruiz, and @fjpedrosa on GitHub. Same person. What I kept from the physics PhD is the best part: the method. I've spent 15 years applying it to products in production and today, as Head of AI Agent, to AI agent systems. Evenings go to my own projects; the latest is Almirant. Still allergic to buzzwords.
// profile.ts
const javi = {
name: "F. Javier Pedrosa, PhD",
aka: "Francisco Javier Pedrosa Ruiz",
phd: "Materials physics · UAM 2017",
method: "hypothesis → measure → repeat",
role: "Head of AI Agent",
building: "almirant.ai",
buzzwords: false,
};
export default javi;// where I come from
2011
Industrial Engineering degree, Industrial Electronics, Universidad de Cádiz. Final-year project: a microcontroller-based electrical energy meter.
2017
PhD in Condensed Matter Physics and Nanotechnology, Universidad Autónoma de Madrid. Thesis: Towards ferrite based rare-earth free permanent magnets (PDF).
15+ years
Product, from whiteboard to production. Startups, growing teams and plenty of opinions about microservices nobody needed. From Málaga.
Now
Head of AI Agent by day. Evenings go to Almirant, Flatzer, Anotea and whatever comes next. More personal every time, and more of it written up on the blog.
What I keep from the PhD isn't the magnets. It's the habit of not trusting a nice result until it repeats. That is exactly what I do with agents: measure, evaluate, measure again.
// what I'm up to
This is not a list of services. It's the list of things my evenings and the odd weekend go into, in case we overlap on any of them.
Almirant is the lab: orchestration, evals and cost per token. Almost everything I learn there ends up on the blog.
Experiments with Codex, Claude Code and Cursor, told with numbers rather than opinions. Sometimes they go wrong; I write that up too.
Flatzer, Anotea, Medianonymizer. Some with users, some in the drawer. Every one taught me something I didn't know.
I have opinions about microservices nobody needed and about saying no in time. If you disagree, even better: that's a conversation.
Physics PhD · 15+ years of product · Málaga · EN / ES
If any of this is close to what you're doing, write to me and we'll talk// latest posts

I asked Codex to cancel my annual Medium subscription and request a refund. Five minutes later, the job was done.
A practical case study on using Codex and a Chrome extension to redesign Mercadona’s classic website without replacing its underlying logic.

How I turned mobile UX improvement into an automated loop that measures, tests, and keeps only changes that move the score from 54% to 100%.
Ten practical rules I use with Cursor, Claude Code, and ChatGPT to reduce noise, preserve context, and get cleaner, more maintainable code.
// how every article ends
// projects


Project management for teams that work with agents. Open source and self-hostable, with scheduled runs, webhooks and MCP access. It started as a tool and is turning into something more personal.


AI agents that answer, qualify and hand off sales conversations for businesses that don't want to see the technical complexity. Validating with the first customers.


Privacy-first media anonymization product for documents, images, audio, and video. AI locates sensitive data; deterministic pipelines apply irreversible redaction with auditability.


AI-powered QR menu platform for restaurants. It turns menu photos into digital menus with dish extraction, allergens, translations, QR flows, analytics, and Google Business integrations.
// contact
I answer myself, not fast but I do. One email and that's it; there is no sales call on the other side.