I came in for three months. I stayed for three years.

TROX makes the kind of equipment most of us never think about. Fans, dampers, ventilation units, fire protection systems, sitting quietly in the ceilings of buildings, keeping the air moving and, occasionally, keeping people alive. In 2022, they started their first software product: a control system for monitoring and managing those units in real time.

I joined as a freelance React engineer on a three-month contract. They renewed it, then renewed it again, every quarter for a little over three years. At the end they offered me a full-time role. I turned it down. Nothing personal. I wanted to stay in the freelance shape: more projects, more rooms where the engineering culture is still being decided.

The work split into three rough categories: code, consistency, and people. Most days I touched all three.

On the code side: in 2022, AI tools weren't anywhere near producing code you could put into a real codebase. So the first stretch was the artisanal phase. Every line typed by hand, every pattern argued out in PRs. A lot of the codebase's bones got laid down there, and deep modules with shallow interfaces was the principle behind most of them. A trending module that hid every category of live data (temperatures, airflow, dampers, smoke, occupancy) behind a one-line interface. A React Context layer that pulled scattered API calls into one place. None of these were fireworks. They were the kind of decisions whose value shows up six months later in a clean PR.

On the consistency side. This is the part I care about most, and the part a lot of engineers treat as someone else's job. When I arrived, the same button could be styled four different ways in four different files: styled-components, MUI inline props, emotion, plain CSS. Four paradigms means four sources of truth for one border radius. It wasn't a styling problem, it was a system problem.
So I made the case for one. Tailwind, applied everywhere, with the design's colours, spacing and type scale defined once as tokens in the config and referenced by name from then on. Not #E5484D - bg-alarm. The payoff was that "does this match" stopped being a judgment call and became a lookup. Fewer abstractions, fewer exceptions, one answer.
Every component went into Storybook: each variant, each state, each edge case, before it ever touched a screen. It made the component the unit of review instead of the page, and it gave the designers and me a shared place to look and discuss.
Then the screens, built pixel-perfect from Figma. Pixel-perfect is not the same as blindly faithful, though. Designs drift too: a padding that's 12 in one frame and 16 in the next, a grey that's almost the one from the palette, a state that was never drawn. When I caught those, I didn't reproduce them. I fixed them in the build, flagged them back, and the design got corrected to match. The system moved in both directions, and it got tighter every time.

About a year in, the same instinct pointed at a bigger gap. External agencies couldn't keep up with the design workload, and inconsistency was arriving faster than me or the team could catch it. I made a sustained case for an in-house designer. A few months after I started making noise, one was hired, and design and code finally had a shared vocabulary.
The other cases I made repeatedly: strict TypeScript with types generated from the GraphQL schema. Testing infrastructure clean enough that things were actually testable. Create React App to Vite. ESLint rules that caught real problems.

On the people side: the frontend team was growing, and a chunk of my time went into helping new engineers get up the curve. Pair programming, code review, the "here's the why behind this pattern" conversations that don't fit in a PR comment. A lot of that feedback was visual. That's 2px off, that's the wrong token, that state has no story. Taste transfers slower than syntax. More than one engineer who came in junior left more senior than I have any right to claim credit for. That's the part I'm most quietly proud of.

One prerelease week I'll mention. I was abroad. The system was going into a real environment for the first time and the bugs were doing what bugs do. I wasn't on call, but I'd built half the thing and knew where the bodies were buried, so I stayed in the codebase until the release went out. Not to brag about working late. I'm driven to finish what I start, and that doesn't show up in a metric.

When I joined, the codebase had:

• Loose TypeScript with nothing enforcing strictness
• Four parallel styling approaches, sometimes inside one component
• No tokens, no component library, no place to see a component outside its page
• No testing infrastructure, no real ESLint guardrails
• Create React App, slow builds, no path forward
• A frontend leaning hard on external design, and drifting from it

When I left:

• Strict TypeScript, types generated from GraphQL
• Tailwind with a token layer in the config. Every colour, space and size defined once, used by name- Every component in Storybook, with its variants and states, as the reference for design and engineering both
• Screens pixel-perfect to Figma, with the inconsistencies caught and corrected on both sides
• A coherent design system with an in-house designer behind it
• Testing, linting and Vite in place, and a team with several engineers who'd come in junior and left ready to lead

None of those are individually flashy. The point is they're still there. Architecture decisions are good when nothing dramatic happens because of them. Design systems are good when nobody has to ask.

TROX was around 20 people on the engineering side by the time I left. Designers, backend, embedded, product, the whole stack. We became friends and family. I miss it.

I also take a clearer picture of what I'm for. I'm a React engineer, but the work I'm best at sits in the seam between design and engineering: taking a direction someone set in Figma and making it hold up as a real system in production. Tokens that mean one thing. Components whose states live in Storybook rather than in someone's memory. Motion that explains instead of decorates.

Freelance is the shape I want my career to be in. More projects, more first-software-product moments, more teams to help scale up. Three years is enough to give to a single team. The rest of my work is ahead of me.