← All work
Cost & Capacity

Right-sizing a fleet with a seasonal load curve

KarpenterCRD controllersFinOps
1,500+ services; 15%+ cost reduction; 1,000+ engineering hours saved.

The problem

Generic Kubernetes cost tooling optimizes toward observed steady state, which is correct for most fleets and wrong for a seasonal business. When your year has a hard peak, the run-up to that peak is precisely when you want headroom that an optimizer reads as waste. Trusting a generic autoscaler through that window means either overriding it constantly or discovering its opinion during your highest-traffic hour.

What I built

Kubernetes-native intelligent autoscaling on CRD-based controllers, with Karpenter handling provisioning and consolidation — tuned for a demand curve generic tooling can't model. Right-sizing where it's safe, deliberate headroom where it isn't, and the policy expressed as a custom resource so teams could reason about it in the same place they reason about everything else.

What was hard

Trust. Autoscaling changes are invisible when they work and career-defining when they don't, so adoption was a change-management problem more than a technical one. Getting 1,500+ services onto it meant making the default conservative and the reasoning legible, not making the algorithm clever.

What this means for vendors

Every vendor in this space demos the steady-state case, and the enterprises with the largest bills are frequently the ones with the least steady-state traffic — retail, tax, ticketing, anything with a season. That gap gets treated as an edge case. It's the opposite: it's where the money is and where generic products are weakest.

Running untrusted agents in production →Platform operations as agent-executable skills →Rules for the 80%, models for the tail →Shipping generative analytics before the playbook existed →Grounding a sales assistant in renewal knowledge →Building the operational view from zero →Scaling content governance with policy and data →