Scalter watches your cases, installed apps and knowledge to surface the next useful action. For that to matter, the recommendation has to land before you finish reading the screen.
We precompute the heavy signals at ingestion time and keep only the light scoring at request time: a handful of features, a compact model, a per-user cache.
Every recommendation carries its reasons. That is what makes it explainable, and what forced us to keep the pipeline simple.
The result: a p95 under 50 ms on our pilot deployments, with no GPU at request time.