Quake Monitor — live earthquake display at Harpa Concert Hall

Iceland has, on a calm day, a half-dozen earthquakes you could plot. The data is public, refreshed every two minutes, and free of charge. Pulling it down is a fifty-line script.

That isn't the project. The project is making an exhibition page that puts each of those earthquakes on the right pixel of a custom-styled Iceland map at 4K — not on a stock web map widget, not on a generic tile layer, but on a hand-rendered satellite image where every coastline, glacier and lava field has been chosen for how it reads on a public-facing exhibition display. And then making the same page look right on an 85-inch screen at Harpa Concert Hall, from a country away, using only photos sent by the installer.

The map is one big static image — a 4K rendering of Iceland, hand-styled for the exhibition. To plot an earthquake on it you need a transformation: real-world (lat, lon) → image (x, y). Get the transformation wrong and a magnitude-three quake appears off the coast of the Westfjords. Get it right and it sits on the volcano it actually came from.

Calibration is the boring kind of hard — pick anchor points, record their (lat, lon) and (x, y), fit the projection, verify, repeat. AI built the tool that turned that into a few minutes of work.

The proof of work is the before-and-after of the projection itself. The first version of the transformation was close, but only close. Many quakes off the south and west coasts plotted into the ocean — visually obvious, technically wrong. AI's QA pass refined the projection until the points sat where they actually came from.

The output isn't a browser. It's an 85-inch 4K panel mounted in a public exhibition, with people standing in front of it every day. That changes the engineering shape. There is no view it on your laptop and ship.

What there also isn't, in this case, is the developer. I never set foot in Harpa, never saw the hardware, never stood in front of the screen. The installer was on-site; I was a country away. The QA channel between the build and the venue was photos.

Two calibration pages were built specifically for that loop. The fit test is a yellow-bordered grid — corners, centre crosshair, gridlines — that makes it obvious whether the entire image is on-screen with no overscan, no letterboxing, no clipped corners. The scale analysis renders a known reference pattern that makes any scaling distortion in the chain visible at a glance. Both pages are designed to be readable from a single phone photo.

The installer took photos of those pages on the actual screen. AI read the photos against expectation, flagged anything that looked off, and once both came back clean the visualisation could ship. The pages exist so a single photo from the venue is enough to QA the whole chain — source image, browser engine, signal, panel — without anyone from the dev team being in the room.

The exhibition page is the public-facing surface. The admin is where the page is operated. Headline text and font, subtext, marker and wave magnitude thresholds, wave scale multiplier, eruption start time, eruption point on the map, colours per element, live preview of the page in miniature — all driven from a single console with a "Save Changes" button that pushes to the live page.

The exhibition page itself is deterministic. AI is not in the loop at runtime — there is no LLM looking at quakes and deciding what to draw. AI's job here was further upstream: build the tooling that made the projection trustworthy, calibrate the page for the physical screen at the venue, and keep the operations console useful.

This is the kind of build that gets dismissed as "just a page" until you have to put it on an 85-inch screen in a public exhibition you've never visited. The interesting work is everywhere except the rendering — in the projection, in the QA loop that runs from installer photos, and in the admin that lets someone other than the developer change a colour without a code deploy.