Build times rot quietly. Nobody notices the day it crosses from fast to slow, they just gradually stop running the full suite locally and start pushing to let CI find out.
We profiled before touching anything and found most of the wall-clock time was spent in steps that rarely changed. Caching those aggressively and only rebuilding what moved did most of the work.
The remaining win was parallelism. Once steps were genuinely independent, running them at the same time turned a coffee break back into a glance at the terminal.