Exercise A: Convert a vague requirement

Requirement:

“Use AI to stop malpractice in online exams.”

Complete:

Business problem:
Decision to improve:
Prediction unit:
Prediction moment:
Target or outcome:
ML task type:
Available features at prediction time:
Model output:
Intended user:
Business action:
Business KPI:
Existing baseline:
False-positive cost:
False-negative cost:
Human approval requirement:

Suggested interpretation

Do not define the target vaguely as “malpractice.”

A more measurable target might be:

Whether an exam session is confirmed by an authorized reviewer
as requiring investigation under defined malpractice rules.

Be careful: a model flag is not proof of malpractice.

Exercise B: AI or normal software?

Choose the primary solution:

  1. Calculate mandatory CSR spending from legally defined inputs.
  2. Rank policies by non-renewal risk.
  3. Retrieve the exact count of active candidates.
  4. Generate a draft explanation of an assessment result.
  5. Disable an account after five failed login attempts.
  6. Identify unusual patterns across proctoring events.

Choose from:

  • C# rules
  • SQL
  • Machine learning
  • Generative AI

Exercise C: Define the baseline

For the exam-risk project, create one simple baseline rule.

Example structure:

Flag the session when:
tab-switch count > approved limit
OR face-missing duration > approved limit.

Then define:

How many sessions does the rule flag?
How many are confirmed by reviewers?
How many confirmed cases does it miss?
How much reviewer time does it consume?

Exercise D: Check feasibility

Answer yes or no:

Do we have historical examples? 
Is the target consistently defined? 
Are the labels trustworthy? 
Are features available before the decision? 
Can someone act on the output? 
Can success be measured? 
Can we run a safe pilot? 
Is the expected benefit worth the cost?
If several answers are no, the project is not ready for model training.
PreviousNext