I’m currently building "LLMParty," a project where I release a new LLM experiment every two days in December (sort of like a coding advent calendar).

Today’s experiment is WanderWorld. It’s an interactive map where you can search for locations using natural language.

The Tech: The core mechanic relies on forcing the LLM to output a specific JSON schema containing coordinates (lat/long). I found that by strictly enforcing the output structure, even smaller, self-hosted models are surprisingly good at geography.

What makes it fun: While it handles standard queries ("Where is Paris?"), it really shines with abstract requests. You can ask for "places that smell like candy" or "best places to see horses" and it maps them out instantly. (don't expect something accurate! :)

I’d love to hear what weird queries you try and if the coordinates hold up!