Public-facing audit and development journal for our AI photo geolocation engine. Every improvement is tracked, every failure is documented. Only the v4.0 post describes the current engine; the earlier posts are archived history.
New here? Start with what geo-estimation is or our guide to finding where a photo was taken.
A vision model, a decision layer backed by 2,265 curated clues, and a 34,000-city gazetteer. 91.2% country accuracy on our 478-photo benchmark of outdoor photographs, in about 9 seconds.
Archived: the December 2025 launch of the GeoLocator dashboard, built on the retired v3.0 engine. For the current engine and its measured results, see the v4.0 post.
Archived: how v3.0 added a reasoning step that explained its evidence before committing to a location. Superseded by GeoLocator 4.0.
Archived: a multi-task model retrained on a non-overlapping validation split after the v1.4 data leak. Superseded by GeoLocator 4.0.
Archived: v1.4 added multi-task learning and border-respecting clustering, but its results were invalidated by an accidental data leak.
Archived: our first baseline classifier, the data and clustering problems that held it back, and what they taught us about geolocation.
Want to see GeoLocator 4.0 on your own photo? Try the free demo.