The pipeline is configurable from the interface itself which helped me a lot on "guesstimating". I modeled my personal understanding after darktable's negadoctor feature. Takes the negative and determines a baseline, and with that come the configurations for color, light. The rolls I work with are sometimes funky and they're tinted differently, for example I have this Lomography Metropolis roll thats a bright neon vomit green. So I also needed a method of manually inputting that, so its captured as density and recorded in a library so numpy can guesstimate the right corrections that need to happen per stock.
Depending how you want to capture it can come from multiple areas, like a blank frame, the trailing* end of developed film, or the gaps in between frames. Interestingly enough along the way I was able to just skip cropping entirely so the script had some extra blank film to go off of for initial estimates.
I just uploaded the code in case you wanted to audit that: https://github.com/shannadige/35mm-studio. I know some others have suggested already-existing libraries that do this, likely better? I've not the time to test it out.
(P.S. for what its worth I found your reply very human, in contrast to most of the wording/phrasing that I let Opus do on my first pass of the blog)
(P.P.S. - OpticFilm 7400 is definitely a good starter scanner. Great output resolution for the price. I see some folks are more partial to camera scanners, but this is easy and quite portable. I got it for 200$ pretty much brand new off of Ebay. I have no intention of upgrading it so I would recommend it. It does need to be paired with good scanning software though, and I think plustek does not maintain any software for Mac. I ended up buying Vuescan a few years back, and it is pretty solid in terms of available controls.)