GetOutCast
A forecasting service that scored the conditions for photography. It was shut down because it cost more to run than it earned.
The problem
GetOutCast was a forecasting service for photographers. It scored forecasts for sixteen kinds of photography, each with hand-tuned scoring weights and an explanation a person could read. Photographers saved their own spots, and could set alerts on them. Running it cost more than it earned.
How I thought about it
Looking back, in money terms: the original cost floor scaled with coverage, not with use. The service pulled NOAA forecast grids on a schedule for every known location, whether or not anyone asked for them. It paid for every forecast up front, for places nobody was looking at.
What I did
The original was built on .NET Framework, with WebForms and MVC. Forecasts were computed ahead of time and served at URLs anyone could list, one after another.
Results
The service was shut down. Its cost model was unsustainable and it generated no revenue. The pre-computed forecasts at listable URLs also made it easy to harvest: it was scraped precisely because the answers were already sitting there. The domain model, the scoring and the explanations, outlived the code.
What I’d do today with AI
The lesson I’d carry forward is money first: cost has to scale with use, never with coverage, and I’d write that down as a rule before any agent writes a line. With a harness I’d make it a check the verify step runs, not a hope. The scoring itself is the part worth keeping, and porting well-understood logic like that is exactly the kind of work agents can do in parallel.