• Build vs. managed platforms

  The AI capability gap between layers has largely closed. The infrastructure gap has not. The right question is where the hard part of your problem actually lives.

• NLU as a tool for the LLM, not the orchestrator

  Moving away from NLU-first architectures was right. Retiring NLU entirely is the overcorrection. The model should orchestrate — the question is what it orchestrates with.

• Eval dimensions and what they're for

  Comprehensiveness feels like rigor. It produces noise. The eval suite should reflect what success looks like for each interaction type, not what the framework can expose.