Elu Gonzalez
@elugon
IT Architect and AI Agent Engineer
I design agents that build, review and verify production software. I set the bar; nothing ships until they clear it.

Behind the calendar: agent branches, pull requests reviewed by CodeRabbit, and tests that decide what lands in each product.
@elugon
I design the system and set the bar. The agents explore, implement, review each other, and only deliver what passes the tests.
Several agents read the repository in parallel: architecture, contracts, tests and debt. Only the questions the code cannot answer reach me, and no line is written until they are closed.
Every change starts with a red test that describes the behavior. Execution agents with a tight scope turn it green. If it passes on the first run, the test is useless.
CodeRabbit reviews every pull request. My resolution agent separates real findings from false positives, puts security and correctness first, and turns failures into regression tests.
The reviewer is never the implementer. For critical deliveries I run independent reviewers in parallel, each with its own angle: integrity, strategy and editorial.
Validators stop a delivery on near-duplicate pages, unsourced data, API spend above the cap, or any failed check in the app's validation manifest.
Each skill lives in its own repository, changes through a reviewed pull request, and passes an audit. It syncs itself at session start. Whatever stops saving work is rebuilt or retired.
Production products I designed, and my agents built and verified with this system.
iOS and watchOS app
App StoreA fitness app with subscriptions, Apple Watch and offline mode. The backend runs on Supabase with RLS, and purchases, App Store notifications and push run in Edge Functions.
150unit and UI test files
How it is verified
An 18-check manifest gates every release, and an AGENTS.md routes agents to 27 architecture policies.
Tax advisory
taxfactory.esBeyond the website, an online contracting system: Stripe payments, document signing and automatic delivery on Workers, D1, R2 and five queues with retries.
328pull requests merged after review
How it is verified
Staging with E2E before every production release, and an AGENTS.md hierarchy checked by a test.
Multilingual content
razfit.appBlog and programmatic SEO for search engines and AI assistants, with llms.txt and agent-ready Markdown on every page.
18languages
How it is verified
Similarity, duplicate and boilerplate detectors, plus a hook that runs the full build and E2E before every push.
Internal tool
My App Store screenshot studio: canvas editor, per-language variants and direct publishing to App Store Connect, on Workers, D1, R2 and queues.
3device formats: iPhone, iPad and Apple Watch
How it is verified
Built with mandatory TDD. If publishing fails halfway, it resumes from the last checkpoint.
My own package
An Astro integration that turns HTML into deterministic Markdown, so an AI agent can read a site without wading through markup.
6client sites use it
How it is verified
CI with coverage and package verification before every release.
Architecture at MSI Studio and my own products, built with agents. Before that, more than ten years in IT consulting.
MSI Studio
September 2024 - PresentFull-timeBarcelona, SpainTeam of 7 developers
I design the architecture and the CI of the studio's products with a team of 7: a European data space, a project and timesheet platform, and an AI rendering product.
Tools: TypeScript, Python, C#, Terraform, AWS, Azure, GitHub Actions
Freelancer
2025 - PresentFreelanceBarcelona, Spain
Separate from MSI, I design products and client platforms and run them with my agents: Cloudflare Workers sites with synced data, programmatic SEO behind validators, and my own apps such as RazFit.
Tools: Cloudflare Workers, D1, R2, Supabase, Astro, TypeScript, Swift
What I build with, and where I have used it.
Most of my agent work lives here. Each skill describes a whole process, like resolving a CodeRabbit review or generating SEO pages, and splits the work between agents that explore and agents that execute. A modified skill has to pass its contract tests before I use it.
I use it to explore large repos and plan before touching code. On sites like Pedro Ochoa I run agents with fixed roles: one for Astro architecture, one for CSS, one for SEO. My skills sync from GitHub every time I start a session.
I am trying it as an orchestrator. On RazFit Web I prepared a plan package so it hands out tasks with the same loop as everything else: explore, test first, separate review.
If you need someone who designs agents and puts them to work on software that holds up in production, write to me.
elugon10@gmail.com