Ran a thirty-person engineering org. Still in the codebase.

Modernizing commerce for the AI era

Forward deployed means I work inside the client's codebase, on their runtime, to their release schedule. The clients are US commerce brands on Shopify Hydrogen. These are headless storefronts, so most of the work is architecture.

Shopping agents are starting to read storefronts and buy from them, not just people. That changes what a commerce platform has to be good at: structured product data, clean APIs, catalog quality, measurement you can defend. It's infrastructure work, and most of the industry is early in it.

Most of the week is the ordinary modernization job. Re-platforming onto Hydrogen, migrating customer accounts and authentication, moving loyalty between providers without losing member state, search and discovery, performance, product information. Nearly all of it ships under live traffic, on stores that are already making money. That constraint shapes every decision.

A lot of the value sits upstream of code: technical discovery, vendor evaluation, and giving a client a defensible reason to pick one architecture over another. On one Hydrogen re-platform I built the QA infrastructure from scratch. Scattered requirements became roughly 585 test cases across eight groups, run across several environments by teams from four organizations. External teams were still using it months after I moved off.

Outside client work I build my own tooling in the open. claude-conductor is a small orchestrator. Claude Code plans the work and decides when it's done, cheaper coding CLIs do the typing headless in isolated git worktrees, and my tests decide what gets merged. About two hundred lines of Python, MIT licensed.

One of twelve, then leading thirty

I joined as the CTO's first hire, one engineer among about twelve in technology, with the CTO running all of it. I didn't inherit a team. I inherited a codebase the previous backend developer had run two refactors through in under a year, leaving it worse than he found it.

I hired most of the backend team before I had the title to lead it. Nobody promoted me into an existing team; the team formed around the hiring and the title caught up. It reached about thirty over that stretch. Head of Engineering came after. I was one of two heads under the CTO, running those thirty across backend, frontend, QA, automation and infrastructure, inside a technology org of about fifty. Data belonged to the other head. All of this while the company went from 35 people to 180, through a Series A led by Quona Capital and into a Series B.

Two systems I conceived and led are still the spine of the product: the commissions and payments module that pays thousands of independent bankers every year, and the document platform that ingests 50,000 documents a month. The real-estate credit flow I built runs R$100M+ a year. Underneath all of it, single server to Kubernetes, thousands of Lambdas, and integrations with Santander, Itaú, Bradesco and a dozen fintechs. Most of them had no real API, so we built one for them.

The people record is the part I'd point to. Two hundred interviews, fifty hires, six engineers grown into leadership roles, including the one who took the job after me. Turnover was among the lowest in the company.

I left slowly and on purpose. Through 2025 I cut my hours and handed the org to the engineer I'd coached into the role, with Form Factory already primary. Formally out in January 2026.

Engineer to team lead inside a year

An international agency handed me a team of ten inside a year. We built a fintech marketplace from scratch that closed $2M+ in deals. We cut 79% off its core process, which ran fifty thousand times a year and thousands of hours of processing. Releases went from three a month to four a week.

I also inherited a client relationship that was going badly and turned it around. Of everything in those two years, that's the skill I've used most since.

Mapping the limits of what a machine can do

My master's was in mechanical engineering, concentration in analysis and mechanical design. The thesis was on the force capability of planar cable-driven robots: 133 pages, written in English, public in the university's repository. I did the work in applied robotics at UFSC's Laboratório de Robótica Prof. Raul Guenther, under Prof. Daniel Martins. The lab runs applied projects with Petrobras, BMW and Celesc, and collaborates with King's College London and Heriot-Watt.

Cable robots can only pull, never push, so the set of active cables changes as the platform moves through its workspace. You can't model the mechanism as a fixed structure. It has to be treated as reconfigurable. I derived those models and defined the laws governing the reconfiguration.

The method combined Davies' method, which applies screw theory and graph theory to statics, with scale-factor analysis and constrained optimization. An intelligent transition system picks the solution approach based on which cables are actuated at that instant. I consolidated it into an algorithm. It produced force capability maps for 2-, 3- and 4-cable robots, and wrench-feasible workspaces for 2- and 3-cable configurations. Benchmarked against the literature it ran more efficiently than methods without the transition, and it guaranteed the platform's pose through exact constraint.

The robots didn't transfer. The habit did: prove where a system holds and where it fails, then show the map. That's most of what it takes to evaluate an AI system honestly. It's also why the sailing piece models a sail as a real wing.

Before UFSC I did a technical degree in electromechanics at CEFET-MG, starting at fourteen, then the engineering degree in mechatronics, robotics and automation at the same institution from seventeen. There was an internship at Hyundai MOBIS in Seoul in 2013.

Partway through that degree I spent a research year at Seoul National University, in Kyu-Jin Cho's biorobotics lab, on a soft meal-assistive exoskeleton for polymyositis patients. It used neoprene bands and two Bowden-cable drives instead of rigid links and joints, and weighed 450 g on the body. The stiffest part of it was a sheet of flexible board. EMG confirmed it worked: the wearer's biceps and deltoid stayed inactive while the device lifted their arm to their mouth. Those cable drives are how I found my master's subject, three years later.

The control interface was mine. Polymyositis patients can't work a joystick or push-buttons, because they have neither the force nor the precision. So I used reed switches. They close on a magnet's proximity, needing no travel and no force at all. Four of them sat on a curved acrylic plate, triggered by a magnet sewn into a thimble, and the patient drove the whole device with the rough wrist motion the disease leaves intact. Published at IROS 2014, the IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 542–547 (PDF). Third of seven authors. I was twenty-one.