From the build log
Stack-neutral deep-dives on answer engines, search, agentic ops, FinOps, and Kubernetes platform work. Patterns first, product second.
Actor/Critic patterns for production remediation (without buying a platform)
Two independent reasoners beat one model that critiques itself. A stack-neutral look at guardrails, calibration, and when not to auto-execute.
Open stack vs proprietary AIOps: compare cost shape, not just unit price
Per-host and per-seat metering compounds differently than inference tokens and shared control planes. A neutral framework for renewal season—no vendor scoreboard.
GPU rightsizing without breaking SLO: condition on p99, not the mean
Average GPU utilization lies. A portable expected-value gate for resizes that keeps availability budgets intact—cloud and scheduler agnostic.
Why one Kubernetes scanner is never enough
Deprecation, policy, reliability, and CVEs are different failure modes. A neutral scoring approach that blends multiple tools without over-fitting weights.
Composable MLOps: bet on glue, not a monolith
Best-of-breed training, registry, serving, and GitOps components age better than a single binary. Trade-offs and when a monolith still wins.