43.36° N · 8.41° W — Galicia, Spain

I put AI agents into production — and measure whether they work.

Agents answering real customers, evals scored against real photographs, and the tooling that got a nine-engineer team shipping this way.

Head of Software Engineering at Innogando · Creator of Trazea

2022: Joined Innogando — 10 people in the whole company
2025: Kubernetes on GKE, GitOps via ArgoCD
2026: Trazea live in both app stores
Now: Team of 9 · AI agents in production

01In production

Slack agent

Natural-language questions over internal databases, answered in Slack. Built with Hermes.

hours → minutesper answered question

WhatsApp agent

Customer support outside business hours, with escalation to a human when the agent is not sure.

real customerslive in production

OCR eval

Accuracy benchmark for a pipeline that extracts expiry dates and supplier lot numbers — scored against real photographs, not clean data.

measuredagainst real photos

Team tooling

A shared Claude Code setup: context files (CLAUDE.md, AGENTS.md), skills, hooks, slash commands and subagents, tuned for how the team actually works.

9engineers using it daily

n8n ops

Lead-classification chatbot (Chatwoot + n8n) and a nightly qualification pipeline running unattended.

400qualified leads

02Selected work

Trazea

2026 — Present

Food traceability and APPCC compliance for hospitality, built end to end with AI tooling: FastAPI + PostgreSQL backend, Expo/React Native apps, authentication, an OCR pipeline, an Astro site, and billing.

Published in both app stores
Favorable technical report from the Food Safety Service of the Xunta de Galicia
Pilot running in real venues

FastAPIPostgreSQLExpoOCRAstro

trazea.es ↗

CoWtrol

Innogando

Internal platform covering stock, orders, support and integrations — one system across sales, assembly and support.

Replaced 5+ disconnected tools

Read the case study →

03Writing

Notes from running AI systems in production — first posts are in the works.

Agent evals

How to know whether an agent actually works — task-level evaluation, not vibes.

Context engineering

Shaping what a model sees: context files, skills, and the tooling around them.

AI in production

What survives contact with real users, and what quietly breaks.

Go to the blog →

04Now

Head of Software Engineering at Innogando — agrotech/IoT, maker of RUMI, GPS collars used by thousands of farmers. I joined when the whole company was ten people; today I lead a nine-person team across product, data and infrastructure, and I still write code.

Infra footnote: Kubernetes on GKE, GitOps via ArgoCD, CI/CD pipelines.