Overview
ai-engineering-from-scratch traces how AI ideas become inspectable, bounded practice: build mechanisms from math, place models in context and contracts, share capabilities, choose workflows or agents, then operate them with evaluation, telemetry, provenance, security, and safety. Its core rule is to let evidence earn each abstraction and limit every claim; toy builds and local checks do not prove mastery or live deployment.
Learning outcomes
- A runnable mechanism makes an abstraction inspectable, but toy or local evidence does not prove production quality, scale, or mastery.
- Context, schemas, retrieval, tools, protocols, and skills should expose information, capability, procedure, and authority as separate boundaries.
- Use the smallest control flow the task justifies: bounded workflows for known paths, agents only when exploration and observable transitions warrant them.
- Durable systems connect outcomes, scope, state, evaluation, telemetry, provenance, security, and safety so missing evidence can narrow the claim.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
What does ai-engineering-from-scratch teach beyond AI concepts?
It teaches a progression from visible mechanisms to bounded, evaluated, and governed AI systems; building and keeping evidence matters more than recognizing terms.
How should a team choose between a workflow and an agent?
Choose a bounded workflow when the path is known; use an agent when exploration is necessary and transitions, authority, budgets, and evidence can be controlled.
Does retrieval, a schema, or a tool make an AI answer trustworthy?
No: retrieval supplies evidence, schemas constrain shape, and tools connect capabilities, but validation, authority, provenance, and evaluation remain necessary.
What does local testing fail to prove about an AI system?
Local tests can establish defined behavior under local assumptions, but they do not prove deployment, transport, scale, security, organizational ownership, or mastery.