Skip to content

🧪 Practice Lab: Quality Architect

Welcome to the architecture office of MagicFridge. Here, we are no longer playing with the chat. We are building the factory. Your mission is to choose the right technological building blocks to solve industrial-scale problems.

Exercise 1: the RAG vs Fine-tuning Dilemma 🏗️

(Objective: choose the right adaptation strategy - LO 4.1.2 vs 4.2.1)

Situation: The marketing team is launching a "Christmas 2026" campaign with very complex discount rules that change every day. GUS (the AI) is completely hallucinating on these rules.

The Project Manager asks you: "Should we retrain the model (Fine-tuning) so it learns these rules?"

Your technical recommendation

What is the best approach?

●
●

Exercise 2: The Secret Agent 🕵️

(Objective: identify an autonomous agent - LO 4.1.3)

Observation: You observe two different behaviors of GUS in the logs. Which one corresponds to the definition of an Autonomous Agent?

You be the judge

Case A: The user asks "Delete my account". GUS replies: "To delete your account, go to Settings > Profile."

Case B: The user asks "Delete my account". GUS connects to the Admin API, checks the balance, executes the SQL DELETE command, and sends a confirmation email.

In which of these cases is GUS an agent?

●
●

Exercise 3: Inside the Database 🛢️

(Objective: understand architectural components - LO 4.1.1)

Context: To make RAG work, MagicFridge must store its thousands of recipes in a way that the AI can "understand" semantically (for example, knowing that "Tomato" is close to "Red Sauce").

Question: What type of database should the architect install?

  1. A Relational Database (SQL)
  2. A Vector Database (Vector DB)
  3. A shared Excel file
See the answer

Answer: 2. A Vector Database.

It allows storing the Embeddings (numerical vectors) of documents. This is what enables the system to perform a semantic search ("find me something that looks like a tomato") rather than an exact keyword search.

Exercise 4: Crisis in Production (LLMOps) 🚨

(Objective: manage operational lifecycle - LO 4.2.2)

Alert: Since this morning's model update, API costs have multiplied by 10, and GUS answers in German half the time.

Which LLMOps practice failed?

  1. Fine-tuning
  2. Monitoring
  3. Orchestration
See the answer

Answer: 2. Monitoring.

A good LLMOps strategy includes automatic alerts on:

  • Cost (to detect abnormal spikes).
  • Quality/Drift (to detect that the AI changes language without reason).

Corrective action: Immediate rollback to the previous model version.


Have you validated the architecture?

If these exercises helped you understand the inner workings, a small coffee for the coach would be greatly appreciated! ☕