Senior Backend & AI Systems Engineer
Java · Quarkus · GraalVM native · Agent infrastructure (MCP, RAG, harnesses) · Next.js
▓▓▒▒░░ · · · · · · · · ░░▒▒▓▓
I'm a senior engineer from Spain. Most of my career has been spent inside Java distributed systems — Capgemini, Accenture, and then TuLotero — and the last couple of years putting agentic AI into production instead of into demos.
I like the unglamorous half of AI engineering: invariants, guardrails, harnesses, regression tests with mutability, and the boring plumbing that decides whether an agent is a product or a party trick. If something I build ends up being useful to someone else, even better — that's why most of it ends up here in the open.
I'm CTO & Tech Lead at Javadaba, where I build the tools I work with every day: a board that people and agents share, and the harness those agents run on. Both are open now, in their Local editions.
░▒▓ ─────────────────────────────────────── ▓▒░
Officium: a backlog you and your AI agents work on together. Agents pick their next card over MCP and move it as they work, review sends work back with its reasons, and what only you can decide waits in a Needs you lane. Free and self-hosted with Docker. |
Fluxor: the harness your agents work with. Rules, skills and hooks kept as versioned items and installed into any repo with |
Both come from FIERCE, my position paper on how to organize software development once agents carry the implementation: a few experts hold the decisions, and nothing counts as done without evidence.
Also open: javadaba-seo-sdk and n8n-nodes-javadaba-seo · fluxor-ide, the desktop IDE where Fluxor started.
Also cooking: RegWatch, a regulatory radar for Spanish normative sources (BOE, AEAT, TGSS, SEPE) on Drools/Kogito · a dithered component library where nothing is an SVG curve and everything is composed from atoms · a pay-per-use API catalog distributed over MCP.
░▒▓ ─────────────────────────────────────── ▓▒░
Selected production work
- TuLotero (15M+ users, high-concurrency distributed system, #1 downloaded app in the Spanish App Store during the Christmas season) — built the public MCP server; ran the pSEO + GEO + MCP initiative directly with the CEO; integrated a new payment provider from scratch with event-based rollback; built the SPEI microservice on Quarkus (~40k transactions/day); distributed + local Redis caching against a WildFly monolith. Proposed the AI workflow adoption that started internal AI use at the company.
- Grupo DIA (Accenture) — 15+ Quarkus microservices in a DDD domain behind a Spring API gateway with caching and multiplexing. Built the payment-methods and customer-addresses services from zero. Saga pattern, Liquibase, Pub/Sub, SQS, SES, BigQuery, testing with Mockito and Karate.
- BMW Group / Alphabet (Capgemini) — vehicle shipment management system; first international Scrum team, ended up coordinating it to hit the deadline.
- Santalucía (Capgemini) — legacy C++/Java microservices against IBM Informix, JSP backoffice, heavy reporting SQL, high L4 ticket volume.
Numbers I can point a commit hash at
- +58% throughput (330 → 520 rps) and p95 295 ms → 162 ms on a production web app
- GraalVM native unlocked: 0.669 s startup, −51% build time
- −98% serverless invocations and −75% compute on an audio fingerprinting pipeline
- ~68,000 programmatic SEO routes shipped with 0–10 ms TBT and CLS 0
- God component refactored 12,077 → 1,802 lines; E2E suite 4m12s → 48s
- Unauthenticated
/mcproute found and closed, documented with a failing test first
These come from an empirical pass over my own git history — traceable to commits, and separated from anything estimated. Plenty of that code was co-authored with coding agents; I don't count it as hand-typed production.
░▒▓ ─────────────────────────────────────── ▓▒░
Languages
Backend & runtime
Spring Batch · EJB · REST · microservices · Domain-Driven Design · API gateway · saga pattern · event-driven architecture · pub/sub
AI & agents
LLM integration · controlled chain of thought · regression testing with mutation · AI-assisted delivery workflows · clustering algorithms
Frontend
Angular / AngularJS · Java JSP
Data & messaging
IBM Informix · Google Pub/Sub · Amazon SQS / SES · Firebase Cloud Messaging
Cloud & DevOps
Lambda · S3 · SQS · SES · Cognito · CDN · App Engine · Cloud Functions · Pub/Sub · Storage · OAuth · CI/CD pipeline design
Quality & observability
Mockito · Karate · quality gates · code review · L3/L4 support · on-call incident response
Payments & integrations
Banking file processing (fixed-position mappers) · programmatic SEO
░▒▓ ─────────────────────────────────────── ▓▒░
Harness engineering as an emerging discipline — the harness as the deterministic partition of the system, while context behaves as a writable partition prone to rot. What a development team is for once one engineer can drive agents through whole epics, which is the question FIERCE tries to answer. Poisoned code: context poisoning, codebase poisoning, adversarial poisoning, and what a real detection posthook looks like. Neuro-symbolic invariant inference. And, further out, what an AI-native OS and its UX should actually feel like.
