Some projects drag on for months. This one didn't.

The idea came from Pascal. He had a clear, sharp intuition: a lot of unstructured text is implicitly spatiotemporal ("a fire near the river last Tuesday"), but downstream systems can't use it because the where and when are never pinned to coordinates and timestamps. If you could extract and index that reliably, you'd have a unified reference layer for retrieval and situational awareness. With his guidance shaping the scope, the rest fell into place fast.

What STIndex does

STIndex is an end-to-end spatiotemporal information extraction system. It uses LLMs to pull locations and times out of free text, resolves them to coordinates and time spans, and exposes the result as a queryable index plus an interactive map dashboard. The point is to give downstream RAG and situational-awareness tools a single, consistent spatiotemporal reference instead of re-parsing raw text every time.

A week, start to finish

What I'm still a little surprised by: from the first sketch to a working system and a written-up paper took roughly a week. Tight scope, a good idea to anchor on, and a lot of momentum. It paid off: STIndex was accepted to the WWW 2026 demo track (Companion proceedings).

Huge thanks to Pascal for the idea and the steady guidance throughout. This one was a genuinely fun sprint.

Try it

There's a live demo at stindex.ai4wa.com. The paper is on Google Scholar.