Part 1 stopped at Verificado19s in 2017. Where the next eight years would land was not something I could have guessed back then. My path might be useful to others, so here’s my perspective.
Globant, again
After Cultura Colectiva I went back to Globant. More seniority this time, and what I’d come back for: leading the AI team. The hook was computer vision for medical imaging. My former boss called me about it, and that was enough.
More models and more deep learning followed. NLP for legal applications, right before ChatGPT. I implemented the end-to-end for a deep neural net for the first time.
Impostor syndrome the whole way through. Some things I got right. Other things I didn’t. One reframe stuck and I still use it: don’t say no. Say “what if we do this instead.”
OPI, then Kavak
I went back to OPI for a bit. While I was there, OPI was acquired by Kavak. I spent the next stretch closing out OPI’s still-active contracts, then went fully into Kavak. Two years there, on the pricing team.
Pricing had two halves: engineering, which made the models work, and data science, which made the models themselves. I led the engineering side. We owned the pipelines feeding Kavak’s pricing across every market it operated in.
It was a jump from consultancy to closing projects to inside a unicorn, stacked close together. Wild. I fumbled more than once. Things stabilized enough that the pipelines did what they needed to do, which is the only bar that matters.
The part worth saying plainly: I was happy and humbled to lead the people I led there.
El Salvador
A friend called and asked if I wanted to move to El Salvador. I said no.
The jobs were stable, the pipelines were working. That’s usually when I start looking around — but moving countries after 16 years in Mexico was harder math. Still: “you always have good ideas, let’s talk.” Two months later, in early 2024, I was changing countries.
When I arrived, my sole responsibility was 5G: mapping tower signal, calculating coverage, optimizing it. Health was a personal objective I brought with me, and I worked it in over time. Other things followed.
The thing I had not felt in this kind of work before was the will to do the right thing pointed in a consistent direction. The intent at the top and the work itself lined up on the same outcome: the people it served. That is not the default. It changes what you can ship.
Years earlier I’d worked inside a national government on AI — the Data Lab at the Mexican Presidency in 2016, a chapter Part 1 didn’t have room for. The texture was similar. The technical problem is usually solvable. The scaffolding around it — political, operational, who-talks-to-whom — is what eats the year. Verificado19s taught me what good decisions during a crisis look like. I wanted that back. El Salvador is small enough that decisions and their effects sit close together.
Then I jumped to health.
Health
This was the type of work I wanted when I came here: a health effort. Someone gets care, gets a prescription, picks up the medicine at a pharmacy. The AI does the parts AI is good at: triage, summarizing the symptoms before the doctor sees them, flagging the things that should not be missed.
I remember launch day plainly. Every AI system in it had been built by me or by the team I led, and the team was brilliant.
The thing I keep from this work is what it felt like to build something people actually used. They opened it on their phones, got a real doctor on the other end, got real medicine, at a scale that’s hard to fathom until you watch the numbers move. And then a next phase, into chronic-disease monitoring.
Shipping into a health system changes the failure modes. On the other end of the API is a doctor with a patient in front of them, deciding whether to trust the system enough to act. That changes what “good enough to ship” means. What you evaluate. What counts as a regression. How much you’ll risk on a Friday deploy.
What the work actually looks like now
The honest version of my day is that I’m a manager who codes. Most mornings I’m in Claude, on the same problems the team is on. Sometimes ahead, sometimes behind, sometimes as a second pair of eyes on something about to ship. The job is to keep the output bar where it needs to be, which mostly means catching what the team didn’t and fixing it on the way past.
This is not how I would have described management to my younger self. Back then I thought the move from manager-who-codes to pure manager was the maturity arc. Let go of the keyboard. Trust the team. Nobody tells you that the keyboard comes back. Claude changed the cost-benefit on staying technical as a manager, so I stayed technical, on purpose. I ship more in a morning than I used to, and the team gets a reviewer who’s already in the code.
What I’d tell my 2024 self
The keyboard came back. Claude made it worth keeping the moment it landed. The teams that ship now are the ones where the senior people are still in the diff. The manager-track-as-departure-from-the-work idea has not aged well.
Will matters more than the model. The technical problem is usually solvable. Whether the system actually ships, gets used, gets believed — that lives in whether intent at the top lines up with intent at the bottom. When it does, the work you can do is not comparable to anywhere else.
The bar is one person opening the thing and being helped by it. Dashboards, metrics, architecture — those exist to serve that one bar. The question is whether someone real, on the other end, used the thing and was better off for it.
That’s the eight years since Part 1.
If any of this is wrong, or there’s a thread you want pulled, reach out.
- #data-science
- #ai
- #llm
- #career
- #el-salvador