Notes organized by the system edges that matter in real products.

Agent workflows become easier to trust when the system can simulate an action before it earns the right to execute it.

The interesting engineering work around AI is often cost control, validation, and fallback behavior.

Approval is not friction for its own sake. It is the product surface that turns a capable agent into an accountable system.

The hard part of billing is rarely the happy path. It is replay safety, idempotency, and keeping state changes boring under stress.

Local products do not become less technical because they serve smaller teams or narrower markets.

Multi-tenant products get safer when organization scoping is visible in code, queries, and docs instead of implied by habit.

The API surface gets easier to trust when response shape, validation, and shared schemas stay more explicit than clever.

Retries, failure reasons, and operator-facing traces matter more than pretending background work is invisible.

A recruiter does not need twenty cards. They need enough evidence to trust your engineering judgment.