2003, and the mouse
The first job was practical computer instruction in São Paulo, plus service calls to homes and small businesses. Students arrived who had never used a computer. In 2003 that was ordinary, and so was watching someone hold the mouse in the air, pointed at the screen, like a remote control.
What surprised me was how little of the work turned out to be about the machine. Explaining a file system to somebody who has never had one is not a technical exercise. It is a translation problem, and the failure mode is always the same. The explanation is correct, complete, and useless, because the person still does not know what to do next. Correct and useless is a category that follows you through an entire career, and 2003 is where I learned to recognize it.
The other half of that job was repair. Somebody’s machine was broken, they needed it for work, and the fix mattered less than whether they understood what had happened and what would stop it happening again. Support work is the same shape at every scale.
The room where the machine is not the point
From December 2009, the work was IT for an aviation-training organization in São Paulo — a regulated one, which changes everything about how a technology decision gets made. Nearly a decade of governance, portfolio delivery, supplier management, budgets and service operations. Scrum, Kanban and SAFe across a multi-year program of change, and a federated governance model so decisions had names attached to them.
Flight simulators came with the territory: operation, maintenance, vendor relationships and the full procurement lifecycle. A simulator is expensive equipment whose entire purpose is to let a person be wrong safely, on the ground, before being wrong in the air. Spending most of a decade around them teaches you something durable about designing systems that expect failure instead of hoping it away.
The other education came from the audit side: internal audit, enterprise risk management, SoX controls, civil-aviation regulatory compliance. Auditors and regulators are a user group. They arrive with a question, and the answer has to be in a form they can act on, or the answer does not count. Same job as the mouse. Different vocabulary, higher stakes, and a hard deadline.
Something else happens when compliance work is normal rather than exceptional. Constraints stop being an obstacle at the end of a project and become an input at the start of one. That habit is the reason guardrails now show up in my designs during discovery instead of 6 months after launch.
Owning the whole thing
From March 2019 the arrangement changed. Sole product and technology owner of Clear Sky Aviation English, an e-learning platform, from concept through to production. Discovery, implementation, vendor management, and then the unglamorous part, which is operating it. There was also a digital workflow built for Germinar RH · OBPI, supporting psychological-assessment experiences used in commercial-pilot evaluation.
Owning everything removes the comfortable idea that the hard part belongs to somebody else. The person who chose the vendor is also the person explaining the outage, and also the person deciding whether the cost is justified. Nothing about that is romantic. It is clarifying. When explanation and consequence sit with the same person, the explanation gets much better, much faster.
The queue
September 2022, Orlando: back to a queue, as a technical support specialist working on endpoint detection and response, antivirus and RMM integration issues for managed service providers. Starting again at the ticket level in a new country is a specific kind of useful. You cannot theorize about how support breaks when you are the one it is breaking on.
A ticket is the purest form of the same problem from 2003. Somebody has a system they cannot act on. Everything else — the tooling, the tiers, the SLAs — is scaffolding around that sentence.
Promotion into support leadership came within 16 months: manager of security technical support. Endpoint security for enterprise and MSP environments worldwide, coaching L2 and L3 specialists, senior escalation point for the situations nobody wanted. Major incident response, root-cause analysis, cross-functional coordination with Product and Engineering, and stakeholders who needed to stay aligned while the ground moved.
Escalations are the exam for this whole discipline: 2 audiences with opposite needs, sharing one conversation. A customer losing money needs to know what happens next and when. An engineering team needs precision and no adjectives. Getting both right, under time pressure, with incomplete information, is explanation as an operational skill rather than a communication skill.
Now the system does the explaining
Since August 2025 the title has been Manager of Artificial Intelligence. Strange thing to arrive at from a computer classroom. Less strange from the queue.
The work is agentic AI across global support and billing operations: stakeholder discovery, workflow design, rollout, evaluation, continuous improvement. LLM-powered virtual agents built with Ada and Mosaic AI inside Zendesk-based workflows, which reached 32% ticket deflection across global support volume. MCP server integrations wired into the Mosaic API and Ada so that agent conversations can be analyzed, failure patterns found, and improvements fed back in. And the security and compliance case that made enterprise adoption possible, which is the document most AI programs never get around to writing.
Designing an agent is teaching. The questions are the ones from 2003, unchanged. What does this person actually need to do next. What can be left out without leaving them stranded. Where does the explanation break, and what happens to them when it does. The difference is that the explanation now runs continuously, without me in the room. That raises the cost of getting it wrong, which is why the evaluation loop sits at the top of the list.
Technology counts when the person on the other end can act on it. Everything else is a demo.
What has not changed
The record now runs 28+ years across technology operations, cybersecurity support, service governance and building things. That is 2 countries, 3 languages, and one recurring conversation with somebody who needs to get something done and cannot.
The AI title arrived in August 2025. Pretending it is older would be a poor start to an essay about explaining things honestly. The instinct behind it is the old part. It was formed by first-time computer users, by auditors who needed an answer they could file, and by customers at the worst point of an outage. None of them cared how the system worked internally. They cared whether they could act.
That is the whole test, and it applies to a support ticket, a simulator procurement, an audit response, and a virtual agent identically. Everything else is production detail.
Written in a personal capacity. Nothing here represents the position of any employer, past or present.