Mainstream map apps expose routing through a handful of toggles — avoid highways, avoid tolls — and a search box that only understands proximity. But the constraints we actually carry in our heads are richer: "keep me off the freeway through downtown," "route around the storm," "grab coffee on the way without backtracking." These are expressible in plain language and computable with open tools, yet they fall outside what mainstream products will do.
This lightning talk shows how to turn natural-language requests into real routes using an entirely open stack: a self-hosted Valhalla routing engine, PostGIS for spatial queries, OpenStreetMap and Overture data, and an LLM that translates intent into routing parameters. The recurring pattern is simple: language becomes constraints, and constraints become geometry.
I'll walk through several examples, route-aware POI search ("coffee on the way"), hazard avoidance ("route around the storm"), and corridor exclusion, then share where it breaks: data coverage gaps, ambiguous language, and the limits of constraint-based routing.