How do you detect slow degradation before full downtime?

Most monitoring tools alert after downtime occurs. In my experience, performance issues usually

start gradually minutes before a full crash.

I built ORVO AI, a lightweight predictive engine that flags instability 60–90 seconds before downtime.

How does your team detect the “in-between” state when a site is technically up but clearly degrading?

Curious to hear approaches from other makers and engineers!