This is the right distinction......Using AI inside a product is very different from owning the AI platform layer. Once you start dealing with RAG quality, provider fallback, evals, tool contracts, observability, HITL, and agent state, it stops being “just add an LLM call.”....The RAGAS numbers are the most useful part here. A system that works in a demo can still have weak retrieval, bad context precision, or hidden failure modes. Measuring that gap is what makes the architecture real.
I also like the build-from-primitives-first approach. Frameworks are useful, but only after you understand what they are hiding.