How geo.qa works.
Earth memory for production AI. The Large Bio-Vision Model forms a 1024-dimensional memory of every place your organisation cares about, scoped per tenant, compatible with the public memory at emem.dev.
A foundation model for places, not pictures.
The Large Bio-Vision Model is a JEPA-based encoder that fuses twelve modalities into one 1024-dimensional memory per place and timestamp. Cameras, drones, satellites, files, OSM, and MCP tools are reduced to a shared vector before they are stored, so every later question runs against one index instead of six.
The 1024 dimensions decompose as 512 visual, 384 semantic, 128 spatial, 128 geo, 64 temporal, and 64 overlay. Memory lands in PostgreSQL with a Vespa vector index next to it. Agents recall from this index before they answer, then cite the cells they used.
Six pipelines, one index.
Every modality terminates in the same Vespa 1024D index. A few shapes shown below; the full set lives on the homepage diagram.
- Satellite pipeline. fetch · spectral · lulc · detect · scene · embed · index. SAR and optical join the same memory cell.
- Camera pipeline. capture · segment · detect · scene · calibrate · embed · index. RTSP frames are geo-registered and scored before they land.
- File pipeline. parse · analyze · segment · context · embed · index. SAM3 handles segmentation; a vision-language model writes the scene caption.
See the homepage for the full six (drone, OSM, MCP rounding out the set).
Two enforcement layers, no outbound writer.
Every memory cell lives under org_{org_id} on the filesystem, and every read is filtered by tenant at the query layer. An agent in tenant A cannot reach a cell in tenant B even when the same address is requested. There is no replication path that crosses the boundary, so the only surface that leaves an org is what a human in that org chooses to share.
Compatible with emem.dev.
emem.dev is the public Earth memory. Apache-2.0, content-addressed by blake3, ed25519-signed, served as cells in the form dedi.zaf00.bafi.baba. It is pre-wired into every geo.qa tenant as a public-facts MCP server, so an agent recalling private memory (1024D) can also cite public bands (1792D) for a 2816-dimensional view of any place.
Receipt means an ed25519-signed claim that a responder asserts a fact; CBOR-canonical, blake3-hashed, verifiable offline. Format shared with emem.dev..