Build the serving MVP · 01
Separate the hot path from safe change
Data plane: request → tokenControl plane: model → deployment
INFERENCE FOUNDER · FIELD COURSE
Build the serving MVP · 02
One model, one GPU shape, one claim
BaselineGatewayMetricsControl loopRollbackReplay
INFERENCE FOUNDER · FIELD COURSE
Build the serving MVP · 03
Autoscaling is late by definition
AdmitRouteDegradeReject
03
INFERENCE FOUNDER · FIELD COURSE
Build the serving MVP · 04
Trace one request before adding one service
Session prefix hitRouter preserves cacheQueue stays boundedVersion is recorded
INFERENCE FOUNDER · FIELD COURSE
Build the serving MVP · 05
Own one decision loop
Observe a workload signalMake a better decisionLearn from the result
INFERENCE FOUNDER · FIELD COURSE