For a workshop on building with AI in 2025, I needed a simple but lively example. I chose a tour guide: share your location with a Telegram bot, walk around town, and it tells you about places nearby. The first version took an hour and a half.
It was very simple inside: coordinates went into a model request, and a story came back. Cursor wrote the code from start to finish. It made a convincing demonstration. For everyday use, it was wonky but interesting.
A walk without a sightseeing checklist
The workshop ended, but the bot stayed. I used it while travelling and close to home. I wanted to find interesting things along the way, without planning a route or following a compulsory itinerary.
Once, an ordinary trip to the post office became a half-day walk. First the bot suggested somewhere nearby, then we wanted to reach the next place. We ended up in a Georgian restaurant that had not even opened yet, where someone poured us chacha. The bot had not recommended the restaurant. At some point, you simply start looking around more closely yourself.
That was the possibility I liked: discovering something where I already happened to be, without having to make a special trip to an “interesting place.”
A story too good to trust
The first version repeated itself, misplaced its map pins and wrote rather dull stories. I added photos, place lookup and detailed editorial instructions: how to connect a story to a place, a person and something you can see for yourself.
I liked one of the facts it sent so much that I became suspicious. I checked it and found inaccuracies and embellishments. The story was interesting enough to make me read more, but not reliable enough to take on trust.
In a 2025 iteration, I connected GPT-5 with web search and rewrote the bot. It gained more thorough checking, French and English interfaces, and a choice of other languages for the stories. The photos still went wrong: sometimes the picture seemed to be telling a story of its own.
Web search added a way to check material against external sources. This describes the 2025 iteration; the later README lists Claude and search-based verification. The models changed as the project developed.
Why it is fine to wait
A walking interface turned out to have an advantage: the answer does not have to arrive immediately. While someone walks and looks around, the bot can spend longer finding a place, checking sources and preparing a story.
So I made live location sharing the main mode. Turn it on, put your phone away, and get an occasional message. A delay that would be irritating in an ordinary chat lets more work happen in the background here. You do not need to stand staring at the screen while AI thinks about your walk.
One day, I went outside and the bot was not working. So I went home to fix it.
A ninety-minute demonstration had become something worth interrupting a walk for. An unusual measure of success for a tour guide.