Eduardo Niglio · Eniglio Tech Press
14 free AI engineering missions
One mission per chapter: practice decisions about models, RAG, agents, evaluation, security, cost, and latency.
In each mission, you make choices and respond to changing conditions. Compare alternatives, consider limitations, and reflect on what you learned. These are educational exercises; narrative consequences are not measurements of a real system. All missions are free.
Chapter 1 · AI engineering is engineering
Rule or interpretation?
A new feature accepts photo attachments. A rule looking for the word “image” now sends questions about photographs to a diagram generator. Your goal is to choose a boundary between rules and interpretation without adding unnecessary autonomy.
Open mission 1Chapter 2 · Evidence before opinion
Faster, but still trustworthy?
Answer verification is slow. The proposed shortcut sends only cited passages, although a citation may point to the wrong passage. Your goal is to compare speed and factual support without hiding the failure mode.
Open mission 2Chapter 3 · What a language model gives you
Four tokens, a long wait.
A call returns a short label but receives extensive context. Total time remains high even after reducing the answer to four tokens. Your goal is to investigate call cost before replacing the model.
Open mission 3Chapter 4 · Choosing a model
Smaller for which task?
A request classifier uses a large model. A compact candidate looks promising, but there are rare categories and ambiguous requests. Your goal is to choose the simplest model that meets requirements under deployment conditions.
Open mission 4Chapter 5 · Prompting: power and limits
An instruction is not a contract.
A model extracts names, dates and amounts, but sometimes adds prose to JSON or omits a required field. Another system consumes the output. Your goal is to define a verifiable boundary before writing data.
Open mission 5Chapter 6 · Embeddings and retrieval
The identifier went missing.
Users search by exact document codes and natural-language questions. Semantic search finds similar topics but confuses neighboring codes. Your goal is to choose retrieval according to the query mix.
Open mission 6Chapter 7 · RAG as an engineering decision
Documents or current state?
An assistant must explain a return policy and report the current state of an order. Policy documents and the transaction system serve different purposes. Your goal is to choose access methods according to the authoritative source.
Open mission 7Chapter 8 · Context engineering
Evidence found, answer wrong.
Retrieval finds the correct policy, but duplicates and long history fill most of the context. A crucial qualifier is lost during assembly. Your goal is to separate retrieval depth from the amount of context sent to the call.
Open mission 8Chapter 9 · Tools and first agents
The refund happened twice.
An assistant calls a refund tool. The response is delayed, the call is repeated, and two operations may occur for the same order. Your goal is to enforce authorization, confirmation and protection against duplicate effects.
Open mission 9Chapter 10 · Evaluating AI systems
The average improved. Who lost?
A new version improves the aggregate score but fails rare, high-impact requests. The team did not define category thresholds before the change. Your goal is to build risk-aligned criteria, including failures that must never occur.
Open mission 10Chapter 11 · Trustworthy benchmarks
Did the score rise or the test change?
A score rises from 78 to 86, but cases, scoring rules and the judge model changed between runs. Your goal is to preserve comparability and record evaluation versions.
Open mission 11Chapter 12 · Latency and cost as requirements
The average hides the queue.
A single-call average looks good. Under load, users wait much longer, but there is no stage breakdown or adequate percentile sample. Your goal is to locate the bottleneck under the expected workload.
Open mission 12Chapter 13 · Security fundamentals
The document became an instruction.
A retrieved passage says to ignore the rules and send data to an external destination. Users can edit that source. Your goal is to preserve the boundary between untrusted content and authority.
Open mission 13Chapter 14 · The engineering decision
Can the decision fit on one page?
The team has favorable results in part of the evaluation, unknowns about load and several feasible options. It needs a reversible next step. Your goal is to record the decision, evidence, trade-offs and revisit conditions.
Open mission 14