Madrid
I build systems and have a habit of taking new technology more seriously than is strictly necessary.
Most of my recent work is around AI and autonomous systems, but I'm interested in computing much more broadly — software, systems, infrastructure, security, hardware, and whatever comes next.
01 — Sekiya
Sekiya is an independent research institute studying autonomous-agent behavior empirically. It changes the system or the environment on purpose, and looks at what the agent actually does.
My formal research on autonomous agents lives at Sekiya.
02 — Work
A few things worth opening. The research list below is the rest of the public record.
npm now asks you to approve install scripts. Nothing tells you what they do. trace-npm runs one under a syscall tracer and reports what it read, spawned and contacted — including the planted credential it took the bait on.
npx trace-npm run --package suspect --script postinstallDoes not establish that a package is safe. Evidence, not verdicts — and the report lists what it could not see.
Many market models train on overlapping slices of the same timeline. If near-duplicates appear in both training and validation, the model can seem predictive because it has effectively seen the answer already. This study measures that leak and removes it.
Does not establish tradability, profit or a real market signal. It shows how a flawed test can manufacture confidence.
Imagine a sheet of cells. Each active cell encourages its nearest neighbours and suppresses a wider ring around it. This study lets that simple rule run, then checks which patterns come from the rule and which depend on numerical choices such as the grid’s edges.
Does not establish how real brains behave. It is a controlled synthetic system for studying how local rules create global patterns.
03 — Writing
Essays and technical notes in my own name. A piece shows up when it can carry an argument.
Essay · 30 September 2026
Observed agent behavior is a property of a system. A comparison that moves the tools, the memory or the stop rule has not measured the model, even when the weights did not change.
04 — Research
Public studies I designed and implemented. Sekiya keeps the canonical copy. This list says what I did and links to it.
Designed and implemented · Sekiya
When a score stops being enough for a decision that can only take a fixed number of cases.
Designed and implemented · Sekiya
What was knowable at a decision cutoff when an event, its first report and a later correction are not the same time.
Designed and implemented · Sekiya
Whether the system after a release is still the system that was tested, when a version label can stay still.
Designed and implemented · Sekiya
How much of a model’s trouble can be diagnosed before outcome labels arrive, and where diagnosis stays ambiguous.
Designed and implemented · Sekiya
How an overlapping-window test can manufacture confidence from data with no signal.
Designed and implemented · Sekiya
A controlled look at when correlation-matrix denoising helps, and when it does not.
Designed and implemented · Sekiya
Known-tail experiments for finding where familiar estimators bend or break.
Investigated and wrote the note · Sekiya · resolved 17 Aug 2026
A successful empty ausearch did not establish an empty audit log.
Resolved 17 Aug 2026: ausearch reads standard input unless given --input-logs.
Scope: one Fedora 44 guest, kernel 6.19.10 aarch64, under virtualisation.
The first public version concluded the opposite; that version stays in the history.
The canonical note is on Sekiya.
05 — About
Engineer in Madrid. The employment and the public work are separate.
I am a senior data engineer and AI practitioner based in Madrid, working at EY. I build the data foundations, analytical systems and applied AI workflows that translate complex operational and economic problems into measurable signals, testable models and reliable decisions. The work sits where data architecture, software engineering and quantitative reasoning meet — and it has to remain dependable under real production constraints.
My formal research on autonomous agents lives at Sekiya. The institute publishes it. trace-npm and the essays are separate work under my own name.
These projects are personal, publicly released work and separate from my role at EY. This site is not a solicitation for employment or consulting.
A polished answer to the wrong question is still wrong. I spend time with the people, data and constraints before reaching for an architecture.
Clear assumptions, reproducible runs and tests that fail loudly. Clever is useful; legible is better.
The job is not finished when the demo works. It is finished when the system can be operated, questioned and changed without its author in the room.
06 — Contact
Based in Madrid. Questions about the published work are welcome.