김태오

Once the live loop existed, I started tuning the control behavior instead of only tuning model metrics.This was where Optuna and Ray/RLlib entered the project. I did not want tuning to be invisible. If the reward weights changed, I wanted the dashboard to show the trials. If the RL policy was disabled for a fast run, I wanted the dashboard to say so. If a PPO bootstrap completed, I wanted the ch..

The six-layer orchestrator was the point where the project stopped being only a data/model pipeline and became an actual control-plane experiment.I built it because I was tired of looking at model scores in isolation. A risk score is useful only if something can consume it. A demand estimate is useful only if it can affect efficiency behavior. Queue pressure is useful only if admission control c..