Skip to content

Part 0 — Summary

Part 0 provided the vocabulary and criteria that guide the rest of the book. What was established:

  • Model vs product vs system. Real value appears in the system; the model is one component.
  • AI systems are probabilistic in deterministic environments. Variance must be isolated.
  • AI is sociotechnical. Operational and organizational failures are often as damaging as model bugs.
  • A demo optimizes for impact; production optimizes for stability under constraints. The two paths almost never converge.
  • Enterprise trade-offs exist between quality, cost, latency, security, governance, privacy, maintenance, scalability and auditability. No system maximizes everything.
  • Deterministic control ≠ probabilistic mitigation. This is one of the central ideas of the book and runs through all the following parts.
  • The architectural decision matrix (Ch. 0.8) is the practical reference for choosing between rules, ML, LLM, RAG, fine-tuning, agent, multi-agent or hybrid.
  • The "Hermes Logística" case will be available as a narrative thread to connect concepts to decisions.

Part 1 moves into traditional ML, showing why it remains the most defensible path for a huge class of problems — including within modern architectures with LLMs and agents.