STEP 1 · Study

AI-300 StudyMachine Learning Operations (MLOps / GenAIOps) Engineer

Read the key points for each exam domain, then use “Practice this domain” to answer questions from that domain only. Domains with a larger weight appear more often on the exam.

Design and implement an MLOps infrastructure

15–20%

Learning goals

Configure workspaces, datastores, compute, and identity and access management
Manage data assets, environments and components, and share them across workspaces with registries
Automate environment setup with Bicep, Azure CLI and GitHub Actions, and restrict networking

Key points

A workspace is the management unit for ML assets. A datastore holds connection information to storage; a data asset is a (versionable) reference to that data.
Choose compute by use: compute instances for development, autoscaling clusters for training, serverless with no management, and so on.
Defining environments (dependencies, images) improves reproducibility of training and inference. Components are reusable building blocks of pipelines.
Registries let you share models, environments and components across multiple workspaces (development, production and so on).
IaC: define workspaces with Bicep and Azure CLI and provision automatically with GitHub Actions. For Azure authentication, consider federated credentials (OIDC) rather than long-lived secrets.
Restrict access to production workspaces with private networking and similar, and manage code in Git.

Workplace uses

SalesKeep the order-forecast model’s development and production environments identical with Bicep, preventing “it worked in dev but not in production.”
General AffairsSeparate workspaces per department and keep view and run permissions minimal with RBAC.
AccountingShare credit score model assets through a registry and place only approved versions in production.

Exam objectives (from the exam guide)

Create and manage resources in a Machine Learning workspace
Create and manage assets in a Machine Learning workspace
Implement IaC for Machine Learning

Learning resources

Learn in detail (3 topics with mini quizzes) →Practice this domain (62 questions) →Next domain ↓

Implement machine learning model lifecycle and operations

25–30%

Learning goals

Implement experiment tracking with MLflow, AutoML, hyperparameter tuning and training pipelines
Register and version MLflow models and evaluate them from a responsible AI perspective
Roll out gradually to real-time or batch endpoints and roll back
Monitor data drift and performance and set up retraining and alerts

Key points

Logging parameters, metrics and artifacts with MLflow lets you compare performance across jobs.
AutoML automatically explores candidate algorithms and preprocessing. Hyperparameter tuning is automated with sweep jobs, and an early termination policy stops wasteful trials.
Use distributed training for large models. Make training steps a pipeline so they can be rerun.
Register and version models. Archive versions no longer in use.
Use real-time (online) endpoints when immediate responses are needed and batch endpoints for periodic processing of large data.
Roll out new versions gradually by routing part of the traffic within the same endpoint, and return to the old version if problems occur.
In production, monitor data drift and performance metrics and trigger retraining or alerts when thresholds are exceeded.

Workplace uses

SalesTry a new lead-scoring model on 10% of traffic and switch fully after confirming the win-rate metric does not worsen.
General AffairsCompute inquiry volume forecasts every morning with a batch endpoint and use them for help desk staffing.
AccountingRaise a drift alert when the input data distribution of the expense-fraud model changes; after retraining, an approver still reviews the results.

Exam objectives (from the exam guide)

Orchestrate model training
Implement model registration and versioning
Deploy machine learning models for production environments
Monitor and maintain machine learning models in production

Learning resources

Learn in detail (4 topics with mini quizzes) →Practice this domain (91 questions) →Next domain ↓

Design and implement a GenAIOps infrastructure

20–25%

Learning goals

Configure Foundry resources, projects, managed identities, RBAC and private networking
Choose between serverless API and managed compute and plan capacity including PTU
Compare prompt variants and version them with Git

Key points

Create Foundry resources and projects with IaC (Bicep / Azure CLI) and control access with managed identities and RBAC.
For foundation model deployment, choose by requirements between pay-as-you-go serverless API and managed compute running on dedicated resources.
Consider provisioned throughput (PTU) for production workloads that need high, stable throughput.
Select models by use (quality, latency, cost), pin the version and roll out to production gradually.
Prompts are assets like code. Create variants, compare them with evaluation and version them in Git.

Workplace uses

SalesCreate three prompts for proposal generation, compare them on evaluation data, tag the chosen one in Git and release it.
General AffairsCompany-wide Q&A is heavily used and needs stability, so it uses PTU; department trials start with serverless API.
AccountingConfigure the accounting project on a closed network and allow access only via managed identities.

Exam objectives (from the exam guide)

Implement Foundry environments and platform configuration
Deploy and manage foundation models for production workloads
Implement prompt versioning and management with source control

Learning resources

Learn in detail (3 topics with mini quizzes) →Practice this domain (73 questions) →Next domain ↓

Implement generative AI quality assurance and observability

10–15%

Learning goals

Create evaluation test data and evaluate with metrics such as groundedness, relevance, coherence and fluency
Configure risk and safety evaluations for harmful content and automated evaluation workflows
Monitor production with latency, throughput, token consumption and traces

Key points

Evaluation starts with a test dataset (questions, context, expected answers and so on) and column mapping.
Main quality metrics: groundedness, relevance, coherence and fluency.
Separately from quality, run risk and safety evaluations for harmful content and similar.
Combine built-in and custom metrics in a workflow that runs evaluation automatically on every change.
In production, continuously monitor latency, throughput, response time and costs such as token consumption, and make detailed logs and traces available to investigate failures.

Workplace uses

SalesEvaluate proposal email generation quality automatically every week and revise the prompt when a drop in relevance score is detected.
General AffairsPinpoint why internal Q&A response latency increased by splitting search and model processing in traces.
AccountingVisualize token consumption per department and use it to manage monthly budget limits.

Exam objectives (from the exam guide)

Configure evaluation and validation for generative AI applications and agents
Implement observability for generative AI applications and agents

Learning resources

Learn in detail (2 topics with mini quizzes) →Practice this domain (40 questions) →Next domain ↓

Optimize generative AI systems and model performance

10–15%

Learning goals

Improve RAG by tuning chunk size, similarity thresholds and retrieval strategy
Validate improvements with embedding model selection, hybrid search and A/B testing
Use fine-tuning and synthetic data and manage them from development to production

Key points

RAG accuracy is decided first by retrieval. Tune chunk size, similarity thresholds, number of results and retrieval strategy while evaluating.
In domains with much jargon, choosing and tuning the embedding model helps. Combine hybrid search when keyword matching matters.
Validate improvements with relevance metrics and A/B tests, not by feel.
Use fine-tuning to improve “output format, tone or specialized tasks” that prompts or RAG do not solve. RAG is better for adding up-to-date knowledge.
When training data is short, create synthetic data, but quality and bias checks are essential. Keep monitoring performance after fine-tuning.

Workplace uses

SalesProduct-manual RAG misses many model numbers, so introduce hybrid search and confirm the effect with an A/B test.
General AffairsFix overly long policy Q&A answers by tuning chunk size and the number of results.
AccountingStabilize the fixed-format output of journal descriptions with fine-tuning, and keep the latest tax information updated on the RAG side.

Exam objectives (from the exam guide)

Optimize retrieval-augmented generation (RAG) performance and accuracy
Implement advanced fine-tuning and model customization

Learning resources

Learn in detail (2 topics with mini quizzes) →Practice this domain (45 questions) →Try the mock exam →

When you finish reading

Practice all domains (311 questions)Mock exam (50 questions · 75 min)AI-300 home

Notes (unofficial, original questions, no pass guarantee)

The questions on this site are original practice questions written by this site based on publicly available skills outlines and exam guides. They are not actual exam questions.
Microsoft, Azure, Copilot, Microsoft 365, Dynamics 365 and Power Platform are trademarks of the Microsoft group of companies. This site is not affiliated with or endorsed by Microsoft.
This content is a study aid and does not guarantee a pass. Exam content, format and prerequisites may change without notice; always check the latest information on the official page before taking the exam.
Your accuracy on the mock exam cannot be converted to the official (scaled) score and does not predict a pass or fail.
Progress is saved only in this browser (localStorage) unless you sign in with Google to sync it. You can delete it from the Cookie & storage settings page (Japanese).