I'm reading that if a random note in the sequence has a high velocity you want the next note to be of a similar velocity relative to the time it's been since last one. Is that in the ballpark?
I would have a random trig for the note of the sequence and a separate sine or triangle lfo that's slower than your sequence, going to a vca that's in line with the amp envelope you're using.
That way, if you have 3 random notes (any number really) in the span of 16 steps you always know that the next one is going to be louder or softer depending on the phase of the lfo.
An envelope follower would work too but I don't have much experience with them