Case study · 2026
Live border waits and cameras for the Balkans, built from official sources and refreshed every hour.
Screenshots of the live site, October 22, 2026.
01The problem
On a Balkan road trip, the border is the one part nobody can plan. Queues run from minutes to hours, and they change by direction, by day and by hour, most of all in summer.
The official figures exist, but they are scattered: auto-club cameras in Croatia, border police readings in Serbia, video feeds in North Macedonia, each on its own site, in its own language and format. Bosnia and Herzegovina and Montenegro publish no wait figure at all.
02The approach
Each official source gets its own small Python scraper that reads at the source's own pace, under a named user agent that says who we are. Everything is normalised into one schema, checked by a freshness gate, and rebuilt into a static Astro site that Cloudflare serves from the edge. GitHub Actions runs the whole loop every hour.
> refresh --hourly
repeats every hour · deploy refused if data is corrupt▋
03The product
One page per crossing: the wait in each direction, the live cameras, who published the figure and when. Where a driver at the border reports something different from the official reading, the report leads, because the person in the queue knows more than a number posted once a day.
04The hard parts
01
Every scraper identifies itself with a named bot, reads no faster than the source publishes, and never solves a CAPTCHA or pretends to be a browser. Where a CDN blocks us, we ask the source before working around it. Where a source publishes nothing, the page says so instead of guessing.
02
The hosted build queue stalled four times in the first season. Rather than babysit it, the hourly job now checks how far production is behind and, past two hours, ships the site through a guarded break-glass path.
03
A third of search impressions were for camera queries owned by an auto-club brand. We measured it: ranking first there earned no more clicks than ranking fifteenth. So we stopped chasing it and put the effort into wait-time searches, which do convert.
05The result
first day with Google impressions, one day after the first commit
Google impressions in the first twelve weeks, with 14,099 clicks
impressions on the peak day, August 7, from searchers in 178 countries overall
average Google position across every query, July 28 to October 20
06What we'd do differently
Over 99% of search demand landed on individual crossing pages; route and corridor pages drew a few hundred impressions between them. We would have put every early hour into the crossing pages and built the rest only once a reader asked for it.
07What's next
A native app built with Expo, sharing its core with the site, is in progress. Partner sites can already embed a live crossing widget, and the season ahead is about getting those embeds onto the travel sites drivers already use.
> next --project
Have something that should run by itself?
