Super minor compared to a lot of the stuff here, but when I was a young data scientist I had the job of creating a model to tell our call center sales agents when an ancillary product would be a good fit to upsell to customers, and when not to waste their time. The only problem: integrating it into the sales application was nearly impossible, IT said it was a huge effort, nobody wanted to do it.

So what I did was built a greasemonkey script to watch the DOM as they went through the sales flow and record the values as they were entered. I then built decision tree model of moderate size, exported it to a string, converted that into a big javascript function. Then, if the model said things were looking good, I modified the DOM to insert a little "alert" box on the top of the page. I handled all the state manually in case they navigated away or did things in a funny order. I knew zero JS or web development at the time, so this was SUPER hacky. But it worked! I then manually walked around to sales agent computers and installed the greasemonkey extension/script. I even got IT involved eventually to serve the script from an internal endpoint, allowing for easier updates.

The actual model ended up being just okay, and didn't have a huge impact on actual sales, but the exec team was SUPER impressed with the delivery mechanism. We had a parent company and they loved to brag to their superiors how we had deployed a machine learning model "for zero IT cost". They had me a do a writeup and everything in case someone wanted to copy my revolutionary idea. I'm sure some guy at the HQ took a look at my writeup and got a good laugh out of how incredibly obtuse, insecure, and hacky the whole thing was.

That said, I still think it was a clever solution and even wondered about turning it into some kind of product at one point.