M
Marcel Strelow
Guest
- Thread Author
- #1
When you open Inverza and see a confidence score, a condition card, or a one-line chat answer, that single number or sentence is the output of eight different data streams that were all queried in parallel, blended, sanity-checked against live observations, and quietly cross-referenced with what the model said yesterday and the day before.
The post walks through each source: the global Open-Meteo forecast as the base layer, directional horizon-cloud probes 40 km in each of the four cardinal directions, air quality for haze and aerosol estimation, the marine API for coastal locations, Overpass / OpenStreetMap landcover queries for terrain context, live METAR observations from the nearest airport for ground truth, high-res regional models (HRRR for the US, ICON-D2 for central Europe, ICON-EU as a wider fallback) for the next 24-48 hours, and previous model runs to detect forecast drift.
It also covers the fail-silent principle that holds the whole thing together: every enhancement is wrapped so that if any one source is down or slow, the base forecast still ships - the user never sees an error, just a slightly less confident answer.
Read the full post.
The post walks through each source: the global Open-Meteo forecast as the base layer, directional horizon-cloud probes 40 km in each of the four cardinal directions, air quality for haze and aerosol estimation, the marine API for coastal locations, Overpass / OpenStreetMap landcover queries for terrain context, live METAR observations from the nearest airport for ground truth, high-res regional models (HRRR for the US, ICON-D2 for central Europe, ICON-EU as a wider fallback) for the next 24-48 hours, and previous model runs to detect forecast drift.
It also covers the fail-silent principle that holds the whole thing together: every enhancement is wrapped so that if any one source is down or slow, the base forecast still ships - the user never sees an error, just a slightly less confident answer.
Read the full post.