ACCELERATED REVIEWS / STUDY COLLECTION
E077
Introduction to AI Engineering
Data, evaluation, deployment, and responsible systems
20 WHY APPLY
original questions explained solutions worked examples
Independent course-topic review. Original learning material.
Not affiliated with or endorsed by Western Governors University.
EDITION 1.0 | 29 OCTOBER 2026
,ACCELERATED REVIEWS E077 / INDEPENDENT STUDY GUIDE
Contents
Module 1: From model to engineered system 3
Module 2: Frame the problem and protect the evaluation 4
Module 3: Match the method to the question 4
Module 4: Evaluate the decision, not only the score 5
Module 5: Make the service dependable in production 6
Module 6: Design for responsible use 7
Compact reference 8
Practice questions 8
Solutions and explanations 9
Three-session study plan 10
Twelve-term glossary 11
Primary references 12
Edition 1.0 - October 2026 2
,ACCELERATED REVIEWS E077 / INDEPENDENT STUDY GUIDE
An independent study companion covering dependable AI systems, data, evaluation, and responsible use. It is not
affiliated with WGU and does not claim to cover every assessment topic. The E077 description emphasizes
programming, mathematics, applied engineering, scalability, reliability, and responsible use; the sections below
expand those themes with original examples. [1]
Module 1: From model to engineered system
Explanation
An AI model maps an input to an output. An AI system includes everything that makes the output useful: a defined
problem, data collection, preprocessing, model or service selection, software integration, user interface, security
controls, deployment, monitoring, and a way to correct failures. Training a model is one possible step. An engineer
might instead use a well-tested existing model, a rules engine, or a human workflow if those meet the need with
lower risk and cost. The question is whether the whole service delivers the required outcome under realistic
conditions.
Start with a user decision and work backward. Specify the action, deadline, and cost of error. Translate those
needs into testable requirements: a miss rate, response-time limit at peak load, fallback, and access controls. An
offline score will not reveal a changed input field or an overloaded queue.
Programming gives the system repeatable behavior; mathematics helps describe error and capacity. Data science
often tests hypotheses, while research may develop algorithms. These roles overlap, but AI engineering
emphasizes operation. Its lifecycle is iterative: frame, prepare data, build a baseline, evaluate, deploy, observe,
and improve or retire. [1][5]
Worked scenario
A campus facilities team wants to prioritize reports of broken air conditioners. The action is to send likely
safety-critical reports to a dispatcher within two minutes. A simple keyword rule is the first baseline. An engineer
then tests a classifier on previously reviewed reports, measures missed urgent cases, and checks whether it
works on short mobile messages. The application records a reason code, lets dispatchers override its queue
position, and routes all reports normally if inference fails. The outcome is timely triage with a known miss rate, not
merely a trained classifier.
Common misconception
"A more accurate model automatically makes a better AI product." Accuracy says nothing by itself about response
time, failure handling, privacy, workload, or the value of the decision.
Retrieval prompt
Without looking back, name five parts of the facilities system besides the model. For each, say how failure would
affect the dispatcher.
Edition 1.0 - October 2026 3
E077
Introduction to AI Engineering
Data, evaluation, deployment, and responsible systems
20 WHY APPLY
original questions explained solutions worked examples
Independent course-topic review. Original learning material.
Not affiliated with or endorsed by Western Governors University.
EDITION 1.0 | 29 OCTOBER 2026
,ACCELERATED REVIEWS E077 / INDEPENDENT STUDY GUIDE
Contents
Module 1: From model to engineered system 3
Module 2: Frame the problem and protect the evaluation 4
Module 3: Match the method to the question 4
Module 4: Evaluate the decision, not only the score 5
Module 5: Make the service dependable in production 6
Module 6: Design for responsible use 7
Compact reference 8
Practice questions 8
Solutions and explanations 9
Three-session study plan 10
Twelve-term glossary 11
Primary references 12
Edition 1.0 - October 2026 2
,ACCELERATED REVIEWS E077 / INDEPENDENT STUDY GUIDE
An independent study companion covering dependable AI systems, data, evaluation, and responsible use. It is not
affiliated with WGU and does not claim to cover every assessment topic. The E077 description emphasizes
programming, mathematics, applied engineering, scalability, reliability, and responsible use; the sections below
expand those themes with original examples. [1]
Module 1: From model to engineered system
Explanation
An AI model maps an input to an output. An AI system includes everything that makes the output useful: a defined
problem, data collection, preprocessing, model or service selection, software integration, user interface, security
controls, deployment, monitoring, and a way to correct failures. Training a model is one possible step. An engineer
might instead use a well-tested existing model, a rules engine, or a human workflow if those meet the need with
lower risk and cost. The question is whether the whole service delivers the required outcome under realistic
conditions.
Start with a user decision and work backward. Specify the action, deadline, and cost of error. Translate those
needs into testable requirements: a miss rate, response-time limit at peak load, fallback, and access controls. An
offline score will not reveal a changed input field or an overloaded queue.
Programming gives the system repeatable behavior; mathematics helps describe error and capacity. Data science
often tests hypotheses, while research may develop algorithms. These roles overlap, but AI engineering
emphasizes operation. Its lifecycle is iterative: frame, prepare data, build a baseline, evaluate, deploy, observe,
and improve or retire. [1][5]
Worked scenario
A campus facilities team wants to prioritize reports of broken air conditioners. The action is to send likely
safety-critical reports to a dispatcher within two minutes. A simple keyword rule is the first baseline. An engineer
then tests a classifier on previously reviewed reports, measures missed urgent cases, and checks whether it
works on short mobile messages. The application records a reason code, lets dispatchers override its queue
position, and routes all reports normally if inference fails. The outcome is timely triage with a known miss rate, not
merely a trained classifier.
Common misconception
"A more accurate model automatically makes a better AI product." Accuracy says nothing by itself about response
time, failure handling, privacy, workload, or the value of the decision.
Retrieval prompt
Without looking back, name five parts of the facilities system besides the model. For each, say how failure would
affect the dispatcher.
Edition 1.0 - October 2026 3
