"We're all on a blank canvas, figuring it out together"

There's a thing that happens in computer vision when a model looks at a scene and actually understands what it's seeing. Not just pixels but objects, context, meaning. I've been chasing that feeling my whole career. It's why I left New Zealand. It's why I'm at Sensat.

I spent years in Auckland building traffic compliance software for local government, cameras catching cars in bus lanes. Good work, good team, and I learned a lot. But a few years in, I knew what I wanted next: somewhere fast-moving, working alongside people I could learn from, building things that would push my skills further than I could push them alone.

My wife and I decided to move to London. She found her job first. I started looking, and I was specific about what I wanted: software engineering roles solving problems that hadn't been solved yet, drawing on computer vision, the space I was most familiar with.

Then I found Sensat's job listing. And before I even finished reading it, I went to the website.

There was a 3D scene in the browser. Point clouds, a mesh, highlighted objects: a real construction site rendered and queryable in real time. I'm a software engineer. I know what it takes to build something like that. I sat there for a while just looking at it.

The job was in Sensat Labs -the R&D team- working on state-of-the-art models integrated into a platform that infrastructure companies actually use. Not research for its own sake. Research that ships.

I applied.

Four months in, I've worked on three things.

The first was a search problem. Sensat has a visual search engine that lets you query a scene in natural language: type "pylons" or "swimming pools" and it finds them in aerial imagery. The problem: ask for a large volume of results and the API would crash. Memory issues. The system was computing intersection-over-union across every bounding box in the result set, which blew everything up at scale.

I fixed it using an indexing method that let us retrieve results without running that full comparison. Smaller footprint, no crashes, genuinely scalable. Our tech lead posted about the compute savings afterwards. It was a small thing in the grand scheme, but it was mine, it worked, and it mattered.

The second is what I think about when I wake up.

I can't say much about it yet, but here's the shape of it: imagine you're managing a construction site. You've got hundreds of assets, ongoing risks, and constantly changing conditions. Right now, making sense of all that means manually clicking through the data. What we're building means you just ask. Natural language, straight into the scene. "What are the risks on this site?" "Where are my assets?" "What's changed since last week?"

It's a cascade of agents, one hands off to another, builds on the answer, hands off again, until something genuinely intelligent comes back. We demoed it internally at our last all-hands. The room got it immediately. That's the feeling you want when you've been deep in something for months: watching other people see it for the first time and lean forward.

Getting this into the product properly is what I'm most focused on. It's going to change how people work with geospatial data. I'm certain of that.

The third project I'm in the middle of right now: fixing a real frustration with AI image generation. When customers try to edit aerial imagery using text prompts, "add a marker in the top-left corner", the model hallucinates. It doesn't know what their top-left corner looks like. It just generates something and hopes for the best.

The solution is a canvas tool. You mark exactly where you want the change, and then the AI works from that. Visual, interactive, no professional software skills needed. Adobe Express, but for a construction site from 400 feet up.

At Sensat, people genuinely care: about each other, and about the work.

When I'm building the image enhancement tool, I'm going back and forth with the designer on how it feels, with the product team on what customers actually need, and with engineers on what's possible. That care is why nothing gets a free pass: if something isn't right, people will tell you, because they care enough to be honest about it.

We're a scaling startup, which means there's no thick layer of process between you and the problem. Your work is visible. Your ideas get stress-tested. You're exposed, in the best way.

I don't know exactly what the next four months look like. But I know what I'm building toward: a platform that looks at the physical world and actually understands it. That makes the invisible visible. That answers questions nobody could answer before without a team of people and a week of work.

“A camera that understands a scene. Still magic to me.”

We're not that far away.

‍

Mark is an R&D Engineer at Sensat Labs. He joined Sensat in early 2026 after relocating from New Zealand.