Note that by virtue of the exponential decay, the score can actually go down no matter what.

Yep but in this case it's a "deserved" decrease of the reputation IMO.
Maybe the number of users could also be a factor in your calculation.

I do not like the UA score because it measures somehow the connection to the top witnesses, which is not appropriate for a community.

I have voiced similar concerns to @scipio since it is a very centralized way to start regarding the concentration of VESTS in the hands of the few. However, the way I understand it, it was only used to initialize the algorithm, so the importance of top witnesses should decrease over time (TBC).

But right, if you already have your own community calculation, that works!

For the engagement indicator, I instead use the length of the comment as a seed for the score (which is questionable, but I haven't found any better option so far).

Among the many projects that I want to work on but can t possibly the find the time for it, I was thinking that developing a spam detector shouldn t be too hard using machine learning and SteemPlus to let users provide samples of spammy content.