I am a research software engineer at Terravek Labs, which means my job is the gap between a scientist's prototype and something that can run unattended against real customer data.
The gap is wider than people expect and it is not mostly about code quality. It is about assumptions that were true of the research dataset and are not true in general: that the imagery is cloud-free, that the projection is consistent, that the timestamps are in UTC, that the file exists.
Our approach is to make the research reproducible before we make it fast. Every Labs project has a pinned environment, a versioned dataset reference and a single command that reproduces the headline figure from scratch. If that command does not work on a colleague's machine, the result is not yet a result.
This is unglamorous and it is resisted, reasonably, by scientists who have a deadline. What has made it stick is that we do it with them rather than to them — a Labs project gets research software engineering time as a matter of course, not as a favour.
The open data releases and the reproduction code for our published papers are on the Labs page.