1.94 million archaeological sites
on one offline map

The task

An institute of archaeology needed one place to see every known archaeological site in the world, with a description for each point, working in the field without a connection. Existing sources were scattered, inconsistent and often wrong.

How we approached it

Research first. A team of 60 AI research agents, coordinated by a human, checked 834 candidate sources across the world adversarially: 115 were removed as unreliable, 441 were corrected. Only then did we design the data pipeline. That stage alone used 11.6 million tokens of model work, which is the kind of volume where a senate of agents is the only way.

Data pipeline in Python. Wikidata, OpenStreetMap via Overpass, Pleiades and national registers were merged, deduplicated and geo-tagged with country and epoch offline. The requirement from the institute was simple and hard: every point must have a description. Version 2.9 delivered that for 1,940,996 objects.

The app. Kotlin and Jetpack Compose, MapLibre Native with OpenFreeMap vector tiles packed as PMTiles, a custom search index, layers by site type (castles, shipwrecks, petroglyphs, caves, megaliths, hoards, battlefields, sieges), protection status by national register, historical base maps (Cassini, General Staff, swisstopo), high-resolution satellite layer, azimuth and route navigation with OSRM, field mode for recording.

Releases. Three major versions, each tested on emulator and device by a human: version 3.0 is the release build with a 473 MB APK, a start time around 1.6 seconds, memory around 1.3 to 1.5 GB on device and zero crashes in test runs. Heavy layers were trimmed from the base map to keep it fast.

Why it matters for your project

This is what “data product” means to us: verified sources, a pipeline that can be rerun, and an interface that works where the user actually is, offline in a field. The same approach runs our catalogue and pricing systems for e-commerce.