START WITH THE FOUNDATIONS

AI Engineering Foundations

Define a task, evaluate errors and put controls around model use.

Free · No sign-in required3 lessonsWorked examples & practiceBeginner

0 of 3 lessons checked · Progress is saved in this browser.

  1. 01 / SYSTEM VIEWTask and data
  2. 02 / SYSTEM VIEWModel evaluation
  3. 03 / SYSTEM VIEWHuman decision

This map connects the course ideas; it is not a detailed engineering schematic.

LESSON 1 / 3 · FREE

Task definitions and baselines

Your goal: Choose an observable outcome before choosing a model.

Understand the idea

Define what the system receives, what it must return and who uses that result. A baseline gives a comparison, such as a simple rule or current manual process. Training fit alone does not show performance on new cases, and a generated answer is not verified evidence.

Worked example

A supplied evaluation set has 100 items. A simple rule answers 70 correctly; a model answers 82 correctly. The gain is 12 percentage points on this set, not proof of production readiness.

Try it yourself

A baseline scores 60 percent and a model scores 75 percent. Find the percentage-point improvement.

What should evaluation data represent?

Background reference ↗ · Original teaching text and fictional practice values.

Continue to lesson 2: Precision and false positives →

LESSON 2 / 3 · FREE

Precision and false positives

Your goal: Distinguish accepted predictions from correct predictions.

Understand the idea

Precision is the fraction of positive predictions that are truly positive under the evaluation labels. Recall asks how many actual positives were found. The cost of each error depends on context; one headline metric cannot fully describe an engineering workflow.

Worked example

Of 20 predicted faults, 15 are confirmed by supplied labels. Precision is 15/20 = 75 percent. You cannot calculate recall without knowing the total actual faults.

Try it yourself

Of 40 predicted positives, 30 are correct. Calculate precision as a percentage.

What extra count is needed for recall?

Background reference ↗ · Original teaching text and fictional practice values.

Continue to lesson 3: Controls and human review →

LESSON 3 / 3 · FREE

Controls and human review

Your goal: Separate model suggestions from authorized actions.

Understand the idea

Map data sensitivity, failure consequences and who can approve actions. Test the model and the surrounding workflow, including fallback and monitoring. Human review must be a defined decision with evidence, not merely a label on an automated action. Risk management continues after initial release.

Worked example

A fictional workflow sends 12 of 80 outputs to review based on its documented policy. That is 15 percent of outputs; this proportion is an observation, not an adequate review threshold for every application.

Try it yourself

A workflow reviews 18 of 90 outputs. Find the review percentage.

Who should define action authority and fallback?

Background reference ↗ · Original teaching text and fictional practice values.

Choose your next step →

Finish with evidence of your understanding.

Explain each worked example in your own words, check the units and identify one assumption that could change the result. Then use the linked labs or roadmap to apply the ideas to a broader task.

Practice progress is a study aid, not a professional qualification or proof of field competence.

Continue in AI Engineering

Apply these foundations in an available course, lab or roadmap. The linked resource’s access rules still apply.

Show 27 more available resources

Open the full studio → · Choose another discipline →

About this introductory course

These original lessons use simplified models and explicitly supplied practice values. Calculations teach concepts; they do not select real equipment, establish code compliance or authorize field work. Follow the linked references and advanced resources for deeper study.

Updated · Study progress stays in this browser. There is no certificate or paid gate for these lessons.