Field notes from a production homelab — capacity planning, inference, incident response, migration, and data protection. First-person, measured against real constraints, with the methodology and the numbers shown.
Refit a 48 GB Blackwell node for 70B-class inference, measured it, and kept the 32B as default. The discipline of declining to ship a capability you built.
Read →
Two controlled benchmarks: storage precision buys disk and speed; only compute precision buys serving VRAM — and resolution sets the ceiling.
Read →
A vector-store backup reported success while doing nothing. Refuting the recorded root cause against the live system, fail-closed hardening, and real alerting.
Read →
Migration planning for virtualization estates after Broadcom's VMware acquisition — including GPU-aware workloads and the cost math.
Read →
Retention is not backup. Why third-party Microsoft 365 protection, restore testing, and monitoring matter before the first data-loss incident.
Read →
A first-24-hour response structure: containment, communications, recovery sequencing, and the controls that decide how bad it gets.
Read →