Home / Case study

How I built the skyline

The skyline on the homepage isn't a video or a stock photo. It's a real-time 3D model of downtown Vancouver, built from city open data and running in your browser. This is how it came together, and what I'd do differently.

buildings in the scene
6,832
trees, 1,893 boats
11,442
hero code + data, gzipped
~490 KB
hosting, data and tools
$0

The brief

I wanted the first screen of my site to say what I do: translate between people with a goal and the people who build the thing. A skyline is a good metaphor for that, and Vancouver is home. The constraints were strict. It had to cost nothing to host, run entirely in the browser, stay smooth on a mid-range phone, and still work for anyone who can't or won't run 3D.

I'm a journalist and communicator, not a graphics programmer. I wrote the brief, made the creative and technical decisions, and directed the build with Claude Code. That's what I mean by AI-assisted development: I'm the editor and product owner, and the AI does the typing.

1. Start with real data

The buildings come from the City of Vancouver's open data portal. Its 2015 footprints give accurate outlines but no heights. The 2009 dataset has heights, including the stepped shapes of podiums and towers. OpenStreetMap fills in towers built after both datasets were published. A script in the repo stitches the three together:

  1. If 2009 parts cover most of a 2015 footprint, it's the same building, so use the 2009 stepped shape.
  2. Otherwise use the 2015 outline with a height from OpenStreetMap, or an estimate if nobody has one.
  3. If OpenStreetMap says the site is now much taller, treat it as a redevelopment and use that geometry.
  4. Any OpenStreetMap building that matches no 2015 footprint is new, so add it.

The script runs on my machine, not in your browser. It crops to the view from the Olympic Village seawall, simplifies every outline, and writes one compact data file. The page never calls the city's servers.

The North Shore mountains use real terrain from Canada's national elevation model, and the shoreline, bridge, SkyTrain line, marinas and street grid all come from the same pipeline.

2. Hand-build what data can't describe

Footprints don't know about domes, saucers or glowing rings. BC Place, Science World and Harbour Centre are modelled by hand as simple shapes and placed with their real coordinates, so they read at a glance. Those same landmarks double as the site's navigation: BC Place leads to Experience, Harbour Centre to About, and so on.

3. Make it look like a photograph

I decided early that this should look real, not stylized. It's blue hour, about twenty minutes after sunset. The theme colours, orange and black, come from the lighting and the interface rather than from painting buildings orange. Windows are generated floor by floor, with some lit and some dark. Red lights blink on the tallest towers. The water reflects the city, and there's bloom, fog and a little film grain to finish.

Several choices came from looking at the picture and arguing with it. The camera started 40 metres up. At 100 metres the downtown core and signature towers clear Yaletown, so that's where it sits now.

4. Keep it fast

The result: about 490 KB gzipped for all hero code and data, against a 700 KB budget, and a mobile Lighthouse performance score of 99 on the homepage when I last measured it.

5. Always have a way out

If someone prefers reduced motion, has no WebGL or has JavaScript off, they get the still poster and the normal navigation. There's a visible "Skip to work" link, the landmarks are reachable by keyboard, and there's a plain menu. A cool effect shouldn't be the only way in.

What's still rough

Heights are only as good as the datasets. The City publishes 2022 LiDAR that would give accurate current heights, but its download server blocks scripts, so those tiles have to be fetched by hand in a browser. That's the next upgrade. Some newer towers are also approximated from OpenStreetMap.

What I took from it

The biggest lesson was about translation. Most of the work wasn't code. It was deciding what "realistic" means, choosing which data to trust, setting a performance budget and checking the result against it. Those are the same skills I use between business teams and developers every day.

Data: City of Vancouver Open Data (Open Government Licence – Vancouver), © OpenStreetMap contributors (ODbL), and Natural Resources Canada's elevation model. Full attribution is in the footer. The code is onGitHub.