Work

Intelligent components that sit inside security operations, turning noisy adversarial signal into intelligence a human can act on.

Behavioural-shift detection in networks, distributed systems, vehicles and industrial contexts — defining "normal" carefully enough that the irregular has somewhere to stand out.

How autonomous agents fail, drift and get manipulated — and the guardrails that keep their reasoning aligned until the task is actually done.

A reading community done right: a Spanish-first Goodreads built to make you read more, not scroll more — real shelves you flip spine-to-cover, each title's honest pulse (including who gave up on it, the number Goodreads hides), reading challenges you win by reading, search by an underlined line, and physical-book swaps with readers near you. No ads, no data selling, no Amazon behind it.

Digital menus and QR workflows for hospitality and POS, with quiet, non-intrusive interaction analytics for local businesses.

Taking low-cost consumer drones apart — protocol, firmware, control — and bringing them under custom command. Presented as a talk at 47CON.

Objective, AI-assisted news analysis — flagging political bias, assessing article quality and issuing trust marks. Reached the regional finals of the Santander X Spain Awards.

Turns plain Markdown into self-contained HTML technical reports — cover page, metadata from frontmatter, custom colors and code highlighting. Runs entirely in the browser, no server, no install.

Zero-knowledge sharing on the edge. Content is encrypted client-side with AES-GCM; the key travels in the URL hash and never reaches the server. Optional password, TTL, burn-after-read.

An urban vehicle comparator: race a bike against a car across a city, with live GPS sync, a podium and a PNG export. Usually the bicycle wins.

An organizer for Kindle notes and highlights that turns scattered annotations into reading statistics worth keeping.

A small project for the local cycling community — stickers, keyrings, clothing and more merchandising to come. Coming shortly.

On how large language models and agents open an attack surface that didn't exist before — and what defending it actually looks like.

A hands-on walk through taking apart low-cost drones and the security lessons that fall out of it.

Peer-reviewed work presented at international venues, including the Blockchain and Applications congress.

My master's thesis: a proactive, self-hosted cybersecurity assistant built on open-source LLM agents and advanced RAG. Knowledge distillation and PEFT fine-tune small models to write YARA, Suricata and Sigma detection rules — scored with a purpose-built metric, RuleBench — running privately on consumer hardware, and outperforming GPT-4 and Gemini on real tasks. Published as a chapter (p. 240) in a volume I also co-edited; the fine-tuned model and rule corpus are open on Hugging Face.

Composable, benchmarked skills that extend coding agents: Apple-silicon containerization, safe disk cleanup, self-patching skills, persistent archive memory — with more on the way.