AI-300: Machine Learning Operations Engineer Associate
A condensed guide to Microsoft Exam AI-300, Operationalizing Machine Learning and Generative AI Solutions, written from Microsoft Learn and organised around the published skills list.
The guide is free. An account keeps it in your dashboard, where each module opens in order.
What is inside
Introduction to AI-300
What Exam AI-300 is, the engineer Microsoft wrote it for, the five skill areas and the bullet counts that show where the marks really sit, how a pass is worked out, what changed when it took over from DP-100, what Microsoft has and has not published to prepare with, and why the documentation links on the official study guide send you to the wrong product entirely.
about 23 min
Machine Learning workspace resources
The four resource bullets that open AI-300, read as decisions rather than as steps. Which dependent resources a workspace is stuck with for life, what soft delete keeps and what it throws away, whether a datastore holds a credential or an identity and who can read that credential, what a compute instance and a compute cluster each cost while nobody is using them, why serverless compute is the answer to most training questions now, and what the built-in workspace roles can and cannot actually do.
about 27 min
Machine Learning workspace assets
The four asset bullets in the "Create and manage assets" group. What separates an asset from a resource and why that one distinction explains most of the module, the three data asset types and why a version can never be deleted, curated against custom environments and the hash that decides whether an image is rebuilt, what a component's interface buys you and how reuse is switched off, then registries: when one copy shared across workspaces beats a copy per workspace, and the three things a registry will not do for you.
about 27 min
Infrastructure as code for Machine Learning
"The five infrastructure as code bullets Microsoft measures in AI-300. How a GitHub Actions workflow proves who it is to Azure without anyone storing a password, the federated identity credential that makes it work and the way it fails when it is wrong, where Bicep stops and the ml extension to the Azure CLI starts, the workflow that runs the deployment, then network access in both directions: a private endpoint for who may come in, a workspace managed virtual network for what may go out, with the modes, the rule types, the costs and the decisions you cannot take back. Closes with what belongs in a repository when half the project is data, and what Azure Machine Learning writes onto a job without being asked."
about 26 min
Orchestrating model training
The eight training bullets in AI-300, taught around the one object they all share, the job. What MLflow autologging records and the places it stops, how a command job is written and submitted, what automated machine learning decides for you and what it will not do in distributed mode, sweep jobs with their three sampling algorithms and three early termination policies, when distributed training earns its complexity, how a pipeline job differs from a command job, and how to compare results across dozens of runs without trusting your memory.
about 27 min
Model registration and versioning
The four registration skills Microsoft measures in AI-300. What a registered model actually is and which properties you can never change, why an MLflow model deploys without a scoring script or an environment and what the MLmodel file has to contain for that to hold, how a feature retrieval specification rides in the model folder and what breaks when it does not, what the Responsible AI dashboard produces and the narrow set of models it will accept, then archiving, restoring and the parts of the lifecycle Azure Machine Learning deliberately will not let you do.
about 26 min
Deploying models to production
The three deployment skills Microsoft measures in AI-300. Why an endpoint and a deployment are separate objects and what that separation buys you, the online against batch decision and what each one charges for, the fields inside an online deployment that decide which request fails first, how mini batches and the default deployment work on the batch side, then testing locally and in place, a disciplined order for diagnosing an endpoint that will not come up, and the traffic split, the mirror and the one command that is a rollback.
about 25 min
Monitoring models in production
"The three monitoring bullets Microsoft measures in AI-300. What drift actually compares and which baseline it compares against, why nothing can be measured until the data collector is switched on at deployment time, the monitor as a scheduled Spark job written in YAML and created with az ml schedule, the lookback windows that decide which rows are compared with which, the metric and threshold for each signal and which direction the threshold points, the difference between a model's statistical performance and its endpoint's latency, then how a retraining trigger is wired through Event Grid without retraining on noise."
about 24 min
Foundry environments and platform configuration
The four platform skills Microsoft measures in the first GenAIOps group. What a Foundry resource holds and what a project holds, the three properties you can only set while the resource is being created and the two that move one way only, how Microsoft Entra ID and Azure RBAC divide control plane from data plane, the five Foundry roles and the scope that makes each one work, the managed identity the platform uses to reach your own data, inbound isolation with a private endpoint against outbound isolation with a virtual network, and the Bicep and Azure CLI shapes that build all of it repeatably.
about 28 min
Deploying and managing foundation models
The four foundation model deployment skills Microsoft measures in AI-300. What a serverless API deployment and a managed compute deployment each reserve and each charge for, how the naming changed between the classic and the current Foundry documentation, how to pick a model from a catalogue of ten thousand without trusting a leaderboard too far, the three version upgrade policies and what happens on a retirement date, a rollout order that keeps a way back, and what a provisioned throughput unit actually buys.
about 27 min
Prompt versioning and source control
The three prompt management skills Microsoft measures in AI-300. Why a prompt has to become a named, versioned object before any of the rest works, the components Microsoft says a prompt is made of and the techniques that shift its behaviour, what a variant is in prompt flow and what a comparison between two of them has to hold constant before it means anything, how Foundry agent versions carry the same idea forward with immutable snapshots and a statistical comparison, then what an agent project's azure.yaml declares, the branch and pipeline shape Microsoft publishes, and what belongs in the repository beside the prompt itself.
about 23 min
Evaluating generative AI applications and agents
"The four evaluation bullets Microsoft measures in AI-300. What a test row has to contain before any evaluator can read it, where the rows come from when you have no production traffic, how a data mapping wires a dataset column to an evaluator input and why a target-generated response needs a different mapping from a stored one, what groundedness, relevance, coherence and fluency each actually measure and which metrics cannot run without a reference answer, how the risk and safety evaluators differ from the quality ones in scale, direction and who runs the judge, and where an evaluation run sits in a pipeline."
about 26 min
Observability for generative AI
"The four observability bullets Microsoft measures in AI-300. Why every signal in Foundry lands in one connected Application Insights resource and what happens until you connect it, what the agent Monitor tab actually charts and how a continuous evaluation rule differs from a scheduled one, the latency formula that separates a slow deployment from a long answer and the metric names that go with it, throughput as a capacity question measured in tokens per minute, why token consumption is the cost metric that matters and where it stops agreeing with the bill, then logging through diagnostic settings, tracing through the OpenTelemetry generative AI conventions, and the dedicated table that keeps prompts away from people who should not read them."
about 26 min
Optimising RAG
"The four RAG optimisation bullets Microsoft measures in AI-300. How to tell a retrieval failure from a generation failure before touching anything, the two runtime dials that cost nothing to move, why chunk size is a trade paid for twice, the four levels of ranking and where each one is allowed to act, what Reciprocal Rank Fusion does with the number 60 and why hybrid search wins on product codes and names, choosing an embedding model and the blunt truth about fine-tuning one on Azure, the Foundry evaluators that put a number on retrieval quality, and how an index alias turns a retrieval change into something you can switch back."
about 27 min
Advanced fine-tuning and model customisation
"The four fine-tuning bullets Microsoft measures in AI-300. When changing the weights beats changing the prompt and when it plainly does not, the three customisation methods and how each one wants its data shaped, the grader that turns reinforcement fine-tuning into a specification you have to write, where synthetic data stops helping and starts feeding the model its own reflection, how to read a training curve and a checkpoint, the two safety gates nobody configures, and what changes operationally once an hourly bill is attached to a model with your company's habits baked into it."
about 31 min
Final review and exam day
The revision page for the last week. All fifty-eight skill bullets in one line each, with the space between them weighted the way Microsoft weights the five areas, the scenario cues behind the decisions that are easiest to get wrong, the settings and services you will meet under two names, and what Microsoft does and does not publish about sitting this exam.
about 33 min
How it was written
These study notes summarise Microsoft Learn material for Exam AI-300, which leads to the Microsoft Certified: Machine Learning Operations Engineer Associate certification. It replaced DP-100. Microsoft publishes no learning path for this exam, so the guide is organised around the published skills list instead, one module per sub-skill group, and every bullet on that list belongs to exactly one module.
Every module is written from the current documentation rather than copied from it, and was checked against Microsoft Learn a second time before publishing, by someone other than the writer. Each module says when it was last checked and links the Learn pages it drew on.
This is an advanced exam. Microsoft asks for a data science background, Python, an entry-level understanding of DevOps, GitHub Actions, and Bicep with the Azure CLI. The prose is written so a non-technical reader can follow it, but the guide is honest that it cannot by itself make someone an MLOps engineer.
For the official list of skills measured, see the Microsoft Learn study guide for Exam AI-300.