I review one existing AI feature and give you a clear plan for making it more reliable. The review covers inputs and data access, retrieval and citations, agent tools, evaluation, failure paths, cost, latency, monitoring, and human approvals where relevant. I use representative cases you can share safely and separate model problems from integration problems. You receive a written assessment with prioritized issues, recommended fixes, and an implementation sequence. Production changes are scoped separately. To start, share a brief workflow overview, architecture or repository access if available, and redacted examples. We will agree on the review boundary before work begins.