This talk presents a practical end-to-end pipeline for building an AI map agent with a fully open geospatial stack. Starting from a natural-language request such as “Show me cafes in Hiroshima City,” I demonstrate how TRIDENT, text2geoql-dataset, and core OpenStreetMap services including Overpass API, Nominatim, and Taginfo can be combined to generate grounded, interactive maps without depending on public shared infrastructure for repeated LLM-driven queries.

The talk has three parts. First, I introduce my self-hosted “osm-planet-in-da-house” environment, where planet-scale OSM services run at home and provide rate-limit-free endpoints for experimentation. Second, I explain how text2geoql-dataset synthesizes and validates natural-language-to-Overpass-QL training pairs by resolving place names, checking tag validity, and confirming that queries return real results. Third, I discuss the path toward running the deepest layer of the agent with a tiny local model on Raspberry Pi-class hardware for offline or low-connectivity settings.

Rather than promising magic, I will focus on concrete architecture, lessons learned, practical constraints, and what is realistically possible today with FOSS.