Captured from the live deployment, not a mockup
01 · The problem
Built a six-stage LangGraph.js pipeline, ingest, detect, forecast, analyse, act and audit, that reads simulated cloud, compute and staffing telemetry and turns what it finds into a reallocation plan.
02 · How it works
- 01
Flagged waste against a baseline per resource type rather than one global threshold, marking anything beyond 1.8 standard deviations or under 20% utilisation before the model sees a row.
- 02
Let Gemini draft the plan but not carry it out: routine P3 moves execute on their own, P1 and P2 changes to production capacity wait in a human approval queue, and every action is written to an append-only audit log.
03 · What it cost, and what it returned
Stress-tested the pipeline across six simulated scenarios, from a quiet day to a product launch and a quarter-end cleanup, and fell back to a local engine when the model was unavailable so a rate limit never stalls a run.