GroundIQ

Every building stands on ground somebody guessed at. A geotech drills a handful of holes across a site, draws a 2D cross-section between them, and the line looks just as confident a kilometre from the nearest borehole as it does right next to one. The drawing hides where the data runs out.

GroundIQ is Fugro’s answer: put everything in a 3D ground model, because a 3D model shows where data is missing and a cross-section hides it. That is the product’s opinion, and it is the thing I was hired to make usable. Screens stay private until the platform is public. The decisions do not.

My role

I lead product design for the platform and set the direction for how it talks to users, how it uses AI, and where it is opinionated. I own products and features end to end, from the first sketch to the merged pull request, and I hire and run the design team.

The users are geotechnical engineers. They do not want a black box, and they are right not to. Every automated step in the product has a place where a human can look at what the machine did before it counts.

Review on the way, not in the way

Building a ground model is a chain of automated steps: interpret the soil tests, draw the layers, fill the volume. The first design had a sign-off gate after each one. It looked responsible. On a live test it was a wall of confirmations between an engineer and the model they had already asked for.

The call was to remove the gates and put the evidence beside the process instead. The build is one screen that stays put from start to finish, in three stages instead of five, because the first three steps of the old design always finished together. Each stage shows the plot an engineer would use to check the work: the material interpretation leads with the same classification chart they learned in university, and clicking a material shows how the data was split. When the model is ready, one button opens it. It never jumps you there.

Review on the way, not in the way. Engineers keep the checkpoints they need. They lose the ceremony.

Showing where the model is guessing

The 3D model is the pitch, but a smooth, fully coloured volume is its own kind of lie. Between two soundings 140 metres apart, the model is interpolating, and nothing on screen said so.

I built a confidence mode for the viewer. Voxels far from a real measurement lose saturation and pick up a hatch. Rings around each sounding mark how far its evidence reaches. Click anywhere and a probe shows the confidence at that column, so an engineer can point at the ground under a proposed footing and get a number for how much of the model there is inference.

The engineering constraint shaped it: the model has no per-voxel uncertainty yet, so confidence is derived from distance to the nearest measurement and the demo says so. It is honest about what it does not know, which is the whole point of the product, made visible. It ran as a demo on real project data and set the direction for the production version.

Re-uploading is not an error

Project data arrives as Excel files, all structured differently, often the same file uploaded twice by two people. The original resolution screen treated every duplicate as a conflict: seventy-five identical amber cards, each with the same paragraph of boilerplate, stacked down the page.

Two changes. The list became dense rows that expand one at a time into a plain comparison, “already in GroundIQ” on one side and “your upload” on the other, with the distance between the two coordinates and the depth difference called out. Then the framing changed. After an upload finishes, one dialog says “Some of these are already in GroundIQ” and offers three buttons that each spell out what will happen. Files that look identical go through in bulk. Files with real differences are held back for review.

Amber is now reserved for things that are actually different. The rest is calm.

One toolbar, tools as modes

The 3D workspace had grown a left rail, a top toolbar, and a set of popovers for layers, assets, history, and each calculation. Every tool had found its own home, so nothing had one.

The redesign is a single top bar. Browsing shows view controls and one selector per analysis. Entering a tool swaps the entire bar for a back button and the tool’s name, and the view controls are deliberately unreachable until you leave. A left dock fuses the model switcher with a layers list that is the scene registry: materials, saved sections, calculation results, 3D assets, all in one place. Tool panels live in the right column, every one of them.

It went through three iterations against stress tests before it held. The second iteration split the chrome in two and I reversed it. Fewer places for things to be is worth a redesign of a redesign.

How we work

I joined in 2025. We haven’t used Figma once.

There is no single method. In the first year I designed five different ways, and each one shipped something. A React shell of the platform where walkthroughs replaced PRDs. Claude Design canvases for early exploration. Designing straight in the production codebase and opening the pull request myself. Pencil sketches photographed into Claude Code. And coaching engineers to design in the same loop, then reviewing their work.

The numbers on the production platform, from December 2025 to September 2026:

Pull requests merged	108
Commits on the main branch	188
Engineers’ pull requests reviewed	51
Issues written	171

About four in five of those commits touch the React frontend. One in five touch the Django backend.

Mockups are specs, not pictures. When a design moves from the shell repo into the platform it is rebuilt against the platform’s own components, never pasted in. And the design system is promoted, not designed: a component earns its place in the shared library by being imported from two or more places, and a sweep of five hundred components found that nine in ten had one caller. One caller is not a pattern.

What changed

Project deliveries that run through platform processing went from two months to a week and a half. That is the number the business watches. What it is worth in revenue is not something I can see.

Delivery	Weeks
Before	8
After	1.5

Leading the team

I hire the designers, and I onboard designers and engineers into the same AI-assisted loop: how to brief a model, how to review what it produces, how to know when to stop. New hires get productive in weeks rather than months because the process is written down and the tools are shared.

Tools built along the way

The design repo. Hi-fi React mockups of the platform, served as static files, deployed on GitHub Pages. It looks like the product but has no data restrictions. Anyone in the company can open it, use it, and PR against it. The walkthrough is the spec, the design follows it. It later became Dry Run.

A research synthesis tool. Session notes go into an inbox, an AI pass turns them into findings, and a small server lets Claude query the whole research base from any session. Patterns emerge across sessions without anyone tagging them.

An Autodesk Forma extension. Sits in Forma’s analysis panel next to Sun, Wind and Noise and tells an architect, at concept stage, how difficult and how expensive the ground under their scheme is. Forty-eight commits of mine on that one.

This is the step toward Head of Product for me. I owned features at Motorsport Games. Here I own products, set strategy, and define how a design team operates without design tools.

The hardest problem

Two things keep me up.

Speed vs value. AI lets us develop fast. How do you make sure you’re adding value, not just features? When everything can be built quickly, the discipline is knowing what shouldn’t be built at all.

The generation gap. Geotechs are an older demographic. They work in PDFs that get printed and filed. There are no digital backups. No standard data formats. When we ingest project data, it arrives in Excel files, all structured differently. Every single one. We have to create standards and opinions about how things should be done, while onboarding people who’ve worked one way their entire careers. The interface can be as modern as we like. It still has to be trusted by someone who files paper.