A company has trained a custom vision model to detect safety violations in construction site photos. The model is performing well in testing, and the team wants to deploy it for use by a web application. The web application will send photos and display the detected violations. What is the first step to make the model available to the web application?

Create a new Custom Vision project and upload the web application's images.

Export the model as a TensorFlow model and host it on a local server.

Publish the trained iteration to a prediction endpoint.

Use the training API to get predictions from the web application.

A custom vision model for detecting cracks in concrete has precision of 0.95 and recall of 0.60. The team wants to reduce the number of missed cracks (false negatives) while keeping false positives manageable. What should they do first?

Raise the probability threshold

Lower the probability threshold used for classification

Switch to a different model type

Add more non-crack images to the training set

A team trained a custom vision object detection model to identify defects in metal castings. The evaluation shows a precision of 0.85 and a recall of 0.80 for the 'crack' class. The production line has two requirements: (1) minimize the number of false alarms because each one stops the line for inspection, and (2) catch at least 90% of actual cracks. The team can only retrain the model with additional data. Which action is most likely to achieve both requirements?

Add more images of cracks and non-cracks, focusing on hard examples that are currently misclassified.

Add only more images of non-cracks to improve precision, accepting a drop in recall.

Switch to image classification to simplify the problem and improve both metrics.

Add only more images of cracks to improve recall, accepting a drop in precision.

A company has a custom vision model that must be consumed by an on-premises application with no internet access. They want to use the model for real-time inference. What is the best approach?

Export the model as a Docker container and run it on-premises

Publish the model to a prediction endpoint and call it from the on-premises app

Retrain the model on-premises using the Custom Vision SDK

Use Azure Arc to connect the on-premises app to the cloud endpoint

A developer wants to automate the training of a custom vision model as part of a CI/CD pipeline. They have a folder of labeled images and want to create a new project, upload images, and train the model using Python. What should they use?

The Custom Vision Prediction SDK

The Azure Storage SDK to upload images and then use the portal

Microsoft Certified:Azure AI Engineer Associate

Domain 4Objective 2

Implement Custom Vision Models AI-102 Practice Questions (Page 1)

Part of the Implement computer vision solutions domain, which accounts for 10-15% of the AI-102 exam. Microsoft does not publish an official question count, but from its 100-minute exam (~40–65 total, ~4–10 in this domain), expect 1–3 from this objective — we provide 36 practice questions to prepare you well beyond it. (estimate)

36questions here

8free pages

7concepts

10-15%of the exam

Questions 1–5

1application · medium
A company has trained a custom vision model to detect safety violations in construction site photos. The model is performing well in testing, and the team wants to deploy it for use by a web application. The web application will send photos and display the detected violations. What is the first step to make the model available to the web application?
Select an answer first
2application · medium
A custom vision model for detecting cracks in concrete has precision of 0.95 and recall of 0.60. The team wants to reduce the number of missed cracks (false negatives) while keeping false positives manageable. What should they do first?
Select an answer first
3expert · hard
A team trained a custom vision object detection model to identify defects in metal castings. The evaluation shows a precision of 0.85 and a recall of 0.80 for the 'crack' class. The production line has two requirements: (1) minimize the number of false alarms because each one stops the line for inspection, and (2) catch at least 90% of actual cracks. The team can only retrain the model with additional data. Which action is most likely to achieve both requirements?
Select an answer first
4expert · hard
A company has a custom vision model that must be consumed by an on-premises application with no internet access. They want to use the model for real-time inference. What is the best approach?
Select an answer first
5application · medium
A developer wants to automate the training of a custom vision model as part of a CI/CD pipeline. They have a folder of labeled images and want to create a new project, upload images, and train the model using Python. What should they use?
Select an answer first

Finished these 5 questions?

Review the revealed explanations, or continue through the curriculum.

Next 5 questions →Back to AI-102

Free Basic Practice is a study aid with revealable answers — not a scored exam. Examers.io is independent and not affiliated with or endorsed by Microsoft. “AI-102” is a trademark of its owner, used for identification only.

Implement Custom Vision Models AI-102 Practice Questions (Page 1)

Questions 1–5

A custom vision model for detecting cracks in concrete has precision of 0.95 and recall of 0.60. The team wants to reduce the number of missed cracks (false negatives) while keeping false positives manageable. What should they do first?

A company has a custom vision model that must be consumed by an on-premises application with no internet access. They want to use the model for real-time inference. What is the best approach?

A developer wants to automate the training of a custom vision model as part of a CI/CD pipeline. They have a folder of labeled images and want to create a new project, upload images, and train the model using Python. What should they use?

Keep studying — Implement computer vision solutions