Downlink is the easy part. We spent a year on the four steps between a raw pass and a tile you can read.
Every conversation about revisit time turns into a conversation about how many satellites you have. That is the least interesting variable. We doubled the constellation in 2024 and the median time-to-tile barely moved, because the bottleneck was never the sky.
Where the minutes actually go
A pass lands as a raw strip with imperfect geometry and no radiometric meaning. Between that and something you can put in a model there are four steps, and three of them used to be batch jobs that ran when a queue felt like it.
Orthorectification against a two-metre elevation model is the expensive one. We moved it onto the ingest path and pinned it to the station that received the pass, which removed a cross-region hop that was costing us four minutes on a good day and forty on a bad one.
What we would not do again
For eighteen months we cached at the wrong layer — after masking rather than before it. Every time we shipped a better cloud mask we invalidated the whole cache. Obvious in hindsight, invisible while you are doing it.