Green Ridge campsites are first-come, first-served. Ground truth lives on a clipboard at HQ, not in a reservation API. Campers need planning context before they drive and situational context after they arrive. Overclaiming availability would erode trust faster than no app at all.
The board is authoritative. The app only mirrors it, with visible freshness, within whatever window someone last photographed it.
Uncertainty is the product condition.
Outcomes.
Freshness as a feature
Every site row carries a freshness state. A clearly described last-known state beats false certainty.
Human review is architecture
Clipboard capture runs photo → vision parse → validate → publish or review queue. Low-confidence reads route to a human before they go live.
Solo stack shipped
Design system, PWA, API, vision pipeline, admin tooling, and marketing, with one owner from constraint through field hardening.
Tested, then built
Still shipping, and not on a release calendar. A counted demand scan set the backlog; a replay over the whole photo archive killed a parser that had looked decisive on one sheet.
Production-linked kit
Semantic tokens and component imports pull from the live product, so surfaces stayed aligned through pivots.
What the full write-up covers
- The problem: a clipboard at headquarters is the only record of what’s open
- The scan: one photo of the sheet becomes a list of open sites, played in the page
- The app: the forest in 3D, every site, and how old the news is
- Reading the sheet: the model reads, a person confirms what it isn’t sure of
- The proof: a cheaper model scored 95.7% on one sheet, then failed on real photos
- Out there, one system, and what is still open, with the full detail a click away in each section