Citourfy is a walking audio guide. You pick a neighborhood, start walking, and it narrates the interesting things as you reach them. The tours are not written by hand. A pipeline builds them, and that pipeline is most of the actual product.
Here is roughly how it works, and the one lesson that cost me the most to learn.
Where the content comes from
No single source has everything, so the pipeline stitches a few together. OpenStreetMap, through the Overpass API, is good at finding what exists and where. Wikipedia fills in the history and the stories. Google Places adds the practical things like opening hours and whether a spot is actually worth the walk. Then a language model turns that pile of facts into narration that sounds like a person talking, not a database being read out loud.
Each source is free or cheap on its own. The value is in combining them well, and in the right order.
The lesson: never trust the model about the world
The tempting shortcut is to ask the language model for everything, including where a place is. Do not. Models are confident and often wrong about facts like coordinates, and a walking tour that sends someone across a busy road to a statue that is really two streets away is worse than no tour at all.
So the coordinates come from real map data, and the model only gets to write about places that have already been checked. There is a review step in the pipeline whose entire job is to reject anything the model invented. It felt like overkill until the first time it caught a point of interest that did not exist.
Batch when you can
The other practical thing: generating one tour means generating dozens of little pieces of narration. Doing that one call at a time is slow and expensive. When the pipeline fans out over fifty or more items, it goes through a batch mode instead, which trades a bit of latency for a much smaller bill. For a small product paying its own way, that math matters.
Most of what I learned building Citourfy was not about prompts. It was about treating the model as one unreliable worker inside a larger system, and building the checks that make its output safe to ship.