We're in closed beta right now. Our rule is simple: publish skill, not promises. So the full methodology and the validated figures behind it go to clients under NDA once an engagement actually starts.
Four stages, and each one is versioned and reproducible on its own. Which means a regulator or reinsurer can sit down and replay any figure we've ever issued, start to finish.
Raw radar and optical archives, read directly at source. Atmospheric and geometric corrections happen under our control, not somebody else's black box.
Radar and optical time series get aligned per parcel. Cross-sensor checks catch and reject artefacts that a single instrument would have let through.
Physics-based models are assimilated with observation, and machine learning corrects the residuals afterward. It never replaces the physics.
Every figure ships with its uncertainty, lineage and version attached. APIs and reports expose the same numbers, so there's never two truths floating around.
Below is the core scientific machinery behind each product family, described at the level of capability and principle rather than implementation. The specific algorithms, parameterisations and data sources live in a secure data room and go to clients under NDA.
How much water each crop is actually finding, tracked day by day through the season, demand weighed against supply. The foundations here are ones agronomic science already established.
Drought read on a calibrated, multi-scale basis. Meteorological, soil and groundwater stress get combined into one signal, rather than left as separate percentiles that don't talk to each other.
A physically grounded yield engine, continuously reconciled against what the satellites actually observe. Outputs come as calibrated ranges with honest bounds, so nobody gets handed a single number dressed up as certainty.
Soil-loss risk, worked out on real terrain and real weather, district by district, refreshed each season.
An orbital water-mass signal, brought down from coarse scale to district level under physical constraints. Long-term trend gets separated from seasonal noise, and uncertainty is quantified per pixel rather than glossed over.
Parcel structure — boundaries, strips, management zones — arrives at decision grade. Every biophysical value stays at the sensor's native resolution; none of it gets resampled upward. In short, we sharpen geometry, not physics.
Radar sees structure and moisture straight through cloud cover; optical reads pigment and vigour. The two independent sensors have to agree before any alert fires. When they disagree, that triggers a review, not an output.
Machine learning corrects models. It doesn't replace them. A network trained to learn residuals stays honest even when the season looks like nothing in its training data — because the physics underneath is still doing the heavy lifting.
A number without an error bar is marketing. Every output carries calibrated uncertainty. And the calibration itself gets tested against held-out seasons, so it isn't just an assumption we make once and forget.
When a source is down, we say so. No silent fallbacks, no synthetic filler standing in for missing data. In an audit especially, an honest gap holds up a lot better than a fabricated value ever would.
Every released figure is tied to a model version, input snapshot and code revision. Clients under contract can ask for a replay of any historical output and get back the same number, to the same value, because the pipeline is deterministic. That's what "audit-ready" actually means once you get past the slogan.