Notes from production
What we are learning building inference infrastructure and detection systems. Mostly specifics, occasionally opinions.
Failover that actually fails over
Most multi-cloud inference setups fail over on paper. Here is what breaks when a region really goes away, and the four checks we run before we call a route healthy.
The sandbox is not the boundary
Most teams building models on financial or clinical data think their real boundary is the sandbox. It isn't. The boundary that matters is the one between staging and production.
Open-weight models are an ops problem
Downloading weights is free. Everything after that — capacity, isolation, versioning, cost attribution — is the part that decides whether your platform team says yes.