Composing science.
One model for every point on the map.
Weather stations, grid cells, fields and plants are nodes. Distance, flow and terrain are the edges. Sonora trains foundation models on that graph as it updates, minute by minute, and reads out forecasts at the exact points you care about — not at the nearest airport.
Build, train, deploy.
Any model of the physical world, from a notebook to a running endpoint, on one open platform.
- 01
Build
Write the model against the hub’s loaders: station records, gridded fields, imagery and telemetry, on one time axis with the lag each really has.
- 02
Train
Fit it on the shared record, or fine-tune a pre-trained graph encoder. Every run is scored on held-out observations before it is listed.
- 03
Deploy
Publish the card and the weights. One click turns them into an endpoint at the coordinates that matter to you.
Models, datasets, deployments.
Each with coordinates, a version, and the score it earned in the open.
Model cards
Weights, config and the card that explains them, versioned together. Loads with a client that feels like the one you already use.
Datasets with coordinates
Station records, gridded fields and imagery, published with their location, their units and their real publication lag.
Deployments
Any listed model becomes an endpoint at your points — REST and a stream, carrying the same numbers its card shows.
How it works, one step at a time.
Three sections so far, each built on real or planted data, each checking its own arithmetic as you read.
Timeseries compression
How a window of a series becomes a short vector: the Haar pyramid built by hand on real hours, and what the vector keeps.
Finding leading signals
Which series move first, checked on a world with a planted answer: surprises, lag profiles, a shortlist of thousands, a held-out year.
Geospatial querying
How a place on the map becomes a query of the store, cell by cell, with the arithmetic checked in your browser.
The neighbourhood is the model.
Real-time geospatial data
One-minute station records, kilometre-scale weather fields, satellite passes and market settlements, on one time axis with their true publication lag.
Graph learning
Message passing across stations, grid cells and assets. The neighbourhood is a learned feature, not a lookup to the nearest sensor.
Open foundation models
Pre-trained once over every variable and every place, released with open weights. A small readout turns the shared representation into your forecast.