1. Why is the machine learning workflow described as iterative rather than linear?○ Because models must be retrained every day○ Because training a model routinely reveals problems in the data — missing segments, inconsistent labels — that send the team back to collect different data○ Because business requirements always change mid-project○ Because algorithms must be tested one after another until one works
2. In a typical machine learning project, where does most of the effort go?○ Training and tuning models○ Understanding the problem, finding the data, and cleaning and labelling it○ Selecting which algorithm to use○ Writing the final report
3. What begins when a model is deployed? Select all that apply.☐ Monitoring, because accuracy in the lab is not accuracy in the world☐ Maintenance and retraining, because the world drifts away from the training data☐ Collection of feedback data, especially human corrections of the model's mistakes☐ The end of the project, since the model is now finished
4. What is the deliverable of a data science project?○ A model running in production making automatic predictions○ Insight and recommendations that lead to a human decision being made and implemented○ A cleaned dataset○ A dashboard, in every case
5. Which question is a data science question rather than a machine learning one?○ Which of these 40,000 orders is likely to be fraudulent?○ Why did our returns rate increase 18% last quarter?○ What is the estimated delivery time for each new order?○ Which support tickets should be escalated automatically?
6. What two lenses must every candidate AI project pass?○ Cost and timeline○ Business value and technical feasibility○ Accuracy and speed○ Executive sponsorship and vendor support
7. Why should an organisation's first AI projects be chosen to succeed rather than to impress?○ Because small projects always return more money○ Because the real purpose is to prove the capability, build the data and deployment muscle, and keep failure survivable so the organisation stays willing to try again○ Because ambitious projects are technically impossible○ Because regulators require a phased approach
8. Netflix paid $1 million for a 10.05% improvement to its recommendation algorithm in 2009 and never deployed the winning solution. What is the main lesson?○ Competitions never produce useful results○ Accuracy is not the objective — value delivered in production is; the engineering cost was not justified, the models did not fit production scale, and the business had shifted to streaming○ The winning team cheated on the evaluation metric○ Recommendation systems cannot be improved beyond a certain point
9. Which role is most commonly under-resourced, and is often the real bottleneck?○ Research scientist○ Data engineer○ AI product manager○ Machine learning engineer
10. Why should stakeholders specify the outcome rather than the method?○ Because stakeholders are not permitted to discuss technology○ Because prescribing a technique can forbid a simpler solution that would work better, whereas stating the problem, constraint, success measure and data lets qualified people choose○ Because methods change too frequently to specify○ Because technical teams dislike being given requirements
11. For a fraud model where one transaction in a thousand is fraudulent, why is overall accuracy a poor acceptance criterion?○ Because accuracy cannot be measured on rare events○ Because a model that flags nothing would score 99.9% while being useless; what matters is the balance between frauds caught and legitimate customers wrongly blocked, which is a business decision○ Because fraud labels are never available○ Because accuracy applies only to classification, not detection
12. What is shadow mode, and why is it valuable?○ Running the model on synthetic data before using real data, to protect privacy○ Running the model on real data and recording its predictions without acting on them, so its performance can be compared against what humans actually did — real evidence at almost no risk or cost○ Running two competing models and automatically choosing the better one○ Deploying the model to a small group of customers before a full launch