Nowcast · 4 min
Why nowcasting is a latency problem, not a modelling problem
Below two hours, the sophistication of your model matters less than how fast the answer reaches the person who needs it.
Ensemble Weather AI Research ·
Precipitation nowcasting covers the zero-to-two-hour window. In that range, the physics is largely advection: the rain that will fall on you in twenty minutes already exists, and is visible on radar, somewhere upwind.
That makes the modelling problem tractable and the engineering problem hard. A nowcast that is 3% more accurate but arrives ninety seconds later is worse in every way that matters to someone deciding whether to move an outdoor event indoors.
Our nowcast path is built around this. Radar volume scans are decoded as they land, not on a schedule. The p95 latency target is 150 milliseconds, and we treat a miss as an incident rather than a performance regression.
The same logic applies to flash-flood alerting. The alert fires on catchment response and soil saturation, not on a rainfall threshold, because the threshold that floods a dry catchment is very different from the one that floods a saturated one — and the difference is the whole warning.
Every response carries the spread
Ensemble members, model weights and p10/p50/p90 on every value — including on the free tier.