Part 1 — Summary¶
This part grounded what did not disappear with the arrival of LLMs. Traditional ML remains the cheapest, fastest and most auditable path for a huge class of problems. Almost every serious enterprise system, even those that look "fully LLM-based", has a classifier, an anomaly detector, a reranker or a classic router at some key point.
What was established:
- Data is the main operational risk. Schema, freshness, drift, leakage. The model is only the last link.
- Model metric, offline metric and business metric rarely coincide. Documenting the chain is a safeguard against illusion.
- Feature stores, experiment tracking, model registry are not decoration; they are the foundations of MLOps.
- Monitoring and retraining must be part of the product, not an ad hoc activity.
- Hybrid is the rule. ML + LLM + agent is the defensible enterprise pattern.
Part 2 enters LLMs assuming this foundation — and treating the LLM as a component within a broader system, not as a universal solution.