김태오

The last major phase was about honesty: were the controller decisions actually touching Kubernetes, and was the dashboard explaining the result without overclaiming?The dual-cluster comparison gave me a strong local framework, but it also exposed a subtle gap. Agent A/B/C decisions were being evaluated through the orchestrator path, while the dashboard was reading real Kubernetes metrics. That m..

The local dual-cluster comparison became the most important dashboard phase of the project.Before this, I had an experimental controller and a live dashboard. That was useful, but it was still too easy to fool myself. I needed a baseline. Not a perfect cloud baseline, but at least a local setup where the same external stimulus hit two Kubernetes clusters: one running the experimental Agent A/B/C..

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..

After the six-layer stack existed, the next problem was making it feel alive. Synthetic traces were useful, but I wanted the dashboard to react to Kubernetes state, not just replayed rows.This phase had two jobs. First, I needed trace and synthetic inputs to keep exercising the orchestrator when no cluster was running. Second, I needed a live Kubernetes loop that could collect a snapshot, run th..

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..

Once the data stopped feeling completely suspicious, I moved from the baseline forecaster to the advanced XGBoost track.The advanced track had a different purpose. The baseline forecaster was mostly a sanity instrument. The advanced XGBoost models were meant to become actual inputs to the orchestrator. That meant I cared about more than one prediction horizon, more explicit feature dictionaries,..