Summary

The practice part changed the book's posture from explanation to operation. Model choice, serving gateways, edge deployment, fine-tuning, agents, sandboxes, retrieval, evaluation, observability, release lifecycle, reliability, human oversight, production data, and operating contracts all become decisions a team must make under time, budget, policy, and incident pressure.

The recurring failure mode is the join between components. A model version goes unpinned. A budget runs with nothing to enforce it, a tenant boundary stays vague, a rollback path is missing. A human gate approves the wrong side effect, or an incident never becomes a regression test. The stack becomes operable only when these joins are named as contracts.

AI infrastructure is not a pile of tools. It exists when SLOs, costs, approvals, data flows, incidents, and tenant boundaries are explicit enough for a team to run. What stays open is which operating contracts will become standard practice as model releases, tool surfaces, and reliability expectations keep moving.