Architecture
Guardrails belong at every boundary, not just the front door
Validating user input is the easy quarter of the guardrail problem. The model's own output, what it hands to tools, and what gets rendered afterwards all need the same fail-closed treatment.
Keeping an AI guardrail release honest
For an AI feature with guardrails, releasing it carefully requires two disciplines: coordinating multi-service promotions with evidence at each step, and keeping the validation environment honest enough that human feedback can be trusted.
What AI features force you to specify
When you add an AI layer to a retrieval feature, data decisions that developers used to make by instinct become explicit contract requirements. Live demos are often the first place you find out which ones are missing.
Bad acceptance criteria ship as bugs
Ambiguity in acceptance criteria doesn't disappear when implementation starts. It becomes a decision made without full context. The lineage problem, staleness semantics, product positioning: caught before a sprint begins, they're conversations. Caught after, they're migrations.
The demo is the test
A feature that passes every test can still be broken in ways the test suite is structurally unable to see. Four kinds of blindness that only surfaced once real stakeholders started using an AI booking assistant in a room.
The shape of shared libraries
A shared library is defined by the shape it presents to its dependents. Two cases make that concrete: an icon component used through an undeclared path, and a one-line logging change that quietly broke a numeric assumption.