←Back to Blog
Course · 23 chapters

Applied AI Engineering

Building agents and automation. A path from how LLMs actually work up to shipping safe, evaluated agentic systems, written for working programmers, with no machine-learning math.

Applied AI Engineering
01

Foundations: How LLMs Work

  1. 01

    Inside a Language Model

    What an LLM actually is: next-token prediction, why "just autocomplete" is true but misleading, and the three things a frozen model cannot do for you.

    5 min read
    →
  2. 02

    How Models Generate Text

    Logits to probabilities, temperature and sampling, the prefill/decode phases, and why models hallucinate.

    8 min read
    →
  3. 03

    What Models Take In

    Multimodal input: how an image becomes patches then tokens, why images cost money, and when OCR still beats native vision.

    6 min read
    →
  4. 04

    Reasoning and Test-Time Compute

    Thinking tokens, buying accuracy at inference, and deciding per task when the latency and the bill are worth it.

    6 min read
    →
02

Communicating with Models

  1. 05

    Prompt Engineering

    Message roles, few-shot examples, chain of thought, and decomposing a job into reusable prompts.

    7 min read
    →
  2. 06

    Structured Outputs & Function Calling

    Forcing machine-readable output, JSON Schema validation, and the function-calling seed every agent grows from.

    Covered in Prompt Engineering Is Software Engineering
    →
03

Giving Models Knowledge (RAG)

  1. 07

    RAG Fundamentals

    The knowledge-cutoff problem, chunking, vector stores, and similarity search.

    6 min read
    →
  2. 08

    Advanced Retrieval

    Hybrid search, reranking, metadata filtering, and query rewriting for a real retrieval pipeline.

    Covered in RAG Is a Search Problem Wearing an AI Costume
    →
04

Giving Models Tools

  1. 09

    Tool Use & the Agentic Loop

    Tool schemas, the call to result to continue loop, routing, and orchestrating many tools.

    6 min read
    →
  2. 10

    MCP & Connectors

    What an MCP server actually exposes, the two ways to reach it, the handshake that discovers its tools, and the four things MCP leaves you to solve.

    7 min read
    →
05

Building Single Agents

  1. 11

    Agent Architectures & Reasoning

    The moving parts of an agent, ReAct, plan-and-execute, and self-reflection loops.

    8 min read
    →
  2. 12

    Memory & Context Engineering

    Working vs long-term memory, compression, context engineering, and state persistence.

    Covered in Building a Single Agent Is a Loop With a Goal
    →
06

Multi-Agent Systems

  1. 13

    Coordinating Multiple Agents

    When multi-agent is worth it, supervisor and worker patterns, handoffs, and orchestration plumbing.

    7 min read
    →
07

Reliability & Safety

  1. 14

    Guardrails & Security

    Input and output filtering, prompt injection, jailbreaking, least privilege, and sandboxing.

    8 min read
    →
  2. 15

    Reliability & Error Handling

    Human-in-the-loop approval, retries, fallbacks, circuit breakers, and mid-run recovery.

    Covered in Agents You Can Trust Are Built From Guardrails
    →
  3. 16

    What You Must Not Send

    What leaves your machine on every call: controller duty, data residency, training opt-out, and the copies you create yourself.

    6 min read
    →
08

Evaluation

  1. 17

    Evaluating Agents & LLMs

    Why there is rarely one right answer, eval harnesses, benchmarks, and LLM-as-judge.

    7 min read
    →
09

Production & Capstone

  1. 18

    Caching Is the Cheapest Win

    Prompt caching, why a stable prefix is the whole trick, and the four ordinary habits that silently destroy a hit rate.

    5 min read
    →
  2. 19

    Running Your Own Model

    Open weights, llama.cpp and Ollama, what self-hosting really costs, and the licence that decides whether you may ship it.

    7 min read
    →
  3. 20

    Serving & Performance

    What a serving engine does, how it scales under load, and keeping cost under control.

    Covered in Shipping Agents Is a Level of Its Own
    →
  4. 21

    Observability & Deployment

    Tracing an agent run, monitoring quality over time, and surviving long failure-prone runs.

    8 min read
    →
  5. 22

    Frameworks & Capstone

    The major toolkits and when to reach for each, then assembling everything into one shipped agent.

    Covered in Shipping Agents Is a Level of Its Own
    →
10

Appendix: Vocabulary

  1. 23

    The Words You Will Hear But Never Use

    Training and classic ML vocabulary: recognise it, do not build it. Pretraining, RLHF, LoRA, distillation, overfitting and the rest.

    8 min read
    →