The Architect

I learned to code at nine with a Pascal book in English and a dictionary missing half its pages. My first program calculated days until summer vacation. It was off by 37 days because I didn't know about leap years. At twelve I copied a hacker website from Geocities so badly I left all their HTML comments in the source for six months. Didn't even know what comments were.

Started freelancing at sixteen. PHP, WordPress, jQuery. Then MooTools before jQuery became dominant. Built a lot of terrible websites for clients who wanted things to "pop more." Learned JavaScript properly when React came out. Spent years doing full-stack work for banks and retailers. Systems that handled money and actual users.

Got into AI in 2015 when TensorFlow dropped. The documentation was basically nonexistent. Stack Overflow had maybe three questions, all unanswered. Took 72 hours to run a single example. Built my first chatbot in 2016. Complete disaster. You'd ask it about the weather and it tried to sell you insurance. But every failure taught me something about the gap between conference demos and production reality.

These days I architect AI systems at Atlax 360. Multi-agent platforms, RAG pipelines, document intelligence. Most of my work is fixing things after the prototype phase ends. Workflow engines that email everyone at 3 AM. Document classifiers that label everything as fraud. RAG systems that retrieve everything except what you asked for. The problems are boring: error handling, monitoring, cost controls, guardrails.

I spend a lot of time on context engineering. Not prompt engineering. Actual architecture for how systems maintain context across interactions. Most RAG implementations fail because they treat context as an afterthought. Multi-agent systems work when each agent owns a clear domain and communicates through structured messages. The coordinator doesn't need to know everything, just which agent to ask.

The hard part of AI isn't getting a prototype to work. It's making it reliable enough that you can sleep at night. Most projects fail for boring reasons: chasing the latest model instead of solving the problem, building demos that fall apart under load, forgetting about rate limits and monitoring. The value is in the problem you solve, not the technique you use.

I have a strict 'No Frameworks' policy for high-level logic. I don't use LangChain or LlamaIndex for the core intelligence loop. I write raw prompts and manage state machines explicitly in code. Frameworks are fine for connectors, but delegating your thinking to a library is a mistake. I want to see the raw string hitting the LLM API. The only exception is the Vercel AI SDK, which I use as a low-level primitive for streaming and UI integration in Next.js/React apps. It's the only tool that respects the 'metal'.

A note on tools: I use AI to write code. I use Cursor and Claude Code every day. English is not my first language, so I use AI to check my grammar. I am not ashamed of this. My value is in the architecture and the reliability of the system, not in memorizing syntax or writing perfect prose on the first try. The job is to ship.

After 19 years I still break things to see how they fail. My GitHub has repos named 'definitely-not-production' and 'this-might-work-v3-final-final.' Still reading error messages in languages I barely understand. The errors just cost more now and sometimes they're hallucinated by the AI I'm trying to debug.

I work 100% remote. Before Atlax 360 I consulted for BBVA, Santander, and other enterprise clients. Built everything from game engines to financial systems. I believe the best code is code you don't have to write, which is why I'm interested in AI that reduces complexity instead of adding layers.