Does AI-generated code still need a production-readiness review?
Yes. AI can accelerate implementation, but it does not prove authorization boundaries, payment behaviour, rollback safety, or secret handling in production.
How do you check a payments release before it goes live?
Test against the live provider contract, exercise retries and refunds, and prove the intended idempotency and rollback behaviour before release.
What does a fractional principal architect do?
Owns architecture and high-risk technical decisions a few days a week, with a written decision log that stays with the team.
What should I bring to a 30-minute risk review?
The product, the release you are considering, the team size, and the decision or failure mode that concerns you most.
Our AI feature demos well. Why is production a different story?
A demo proves the happy path once. Production has to survive adversarial input, cost spikes, model updates, and tenant boundaries. We review evaluation coverage, guardrails, fallback behaviour, and data isolation before an AI feature ships.
AI writes a growing share of our code. How do we keep review meaningful?
Treat every AI-written change and every AI review finding as a claim to verify, not a fact to trust. We help teams put deterministic gates before human review, so people spend judgment only where it is actually needed.
How do we stop AI coding agents from leaking secrets or pulling in risky dependencies?
With a written, machine-followable policy: scoped credentials, pinned dependencies, install scripts disabled, and automated leak checks in the pipeline. We wrote ours after a real incident - and it has held since.
Our cloud and AI costs grow faster than usage. What are the architectural levers?
Cost is an architecture property, not a billing surprise: right-sized models and infrastructure, caching, batching, and clear cost ownership per feature - with the invariants written down before anything is cut.
Can we add AI features to a multi-tenant product without leaking data between tenants?
Yes - if isolation is enforced deterministically at every layer the AI touches: retrieval scope, prompts, caches, logs, and telemetry. A language model must never be the thing that enforces access control.
When is a rewrite justified - and when is the boring architecture right?
Rewrite when the current design blocks a measurable business outcome; otherwise evolve it. We write the non-negotiable invariants first, keep a documented way back, and treat 'no change' as a valid recommendation.
