The AI-300 training teaches you how to operationalize machine learning and generative AI solutions on Azure. You start from a model trained in a notebook and take it all the way to a deployed, versioned, monitored and retrainable service. Over four days you work with Azure Machine Learning, Microsoft Foundry and GitHub Actions, using the official Microsoft labs.
The programme follows the full lifecycle: experimentation and model tracking with MLflow, hyperparameter tuning, reusable pipelines, then the design of an MLOps architecture built from the outset for monitoring and retraining. The second half applies the same discipline to generative AI agents: prompt management in GitHub, structured evaluation experiments, automated evaluations, monitoring and tracing of deployed applications.
By the end of the four days you can ship a model or an agent through a reproducible and auditable pipeline, and read production signals to decide when retraining is needed. The course runs in Geneva and Lausanne, on site or as a virtual class, and is delivered by a Microsoft Certified Trainer.
Module 1: Get started with machine learning in Azure
Module 2: Experiment with Azure Machine Learning
Module 3: Run training scripts and track models with MLflow in Azure Machine Learning
Module 4: Perform hyperparameter tuning with Azure Machine Learning
Module 5: Run pipelines in Azure Machine Learning
Module 6: Design a machine learning operations solution (MLOps)
Module 7: Automate model training with GitHub Actions
Module 8: Deploy and monitor a model in Azure Machine Learning
Module 9: Plan and prepare a GenAIOps solution
Module 10: Manage prompts for agents in Microsoft Foundry with GitHub
Module 11: Evaluate and optimize AI agents through structured experiments
Module 12: Automate AI evaluations with Microsoft Foundry and GitHub Actions
Module 13: Monitor your generative AI application
Module 14: Analyze and debug your generative AI app with tracing
Microsoft released the AI-300 course on 20 March 2026. It fills a real gap: until then the official catalogue taught how to design and train models, not how to keep them running. AI-300 is the first Microsoft course fully dedicated to MLOps and GenAIOps, meaning the deployment, monitoring and governance of machine learning models and generative AI solutions. ITTA adds it to the catalogue this autumn in its official form, fourteen modules delivered over four days.
The gap it addresses is familiar to every data team. A model with excellent test scores creates no value until an application calls it, someone watches it, and it gets retrained when the data drifts. The same is true of a generative AI agent: a prompt that shines in a demo degrades as soon as it meets real traffic, and without automated evaluation nobody notices before the users do.
AI-300 is not an entry level course. It assumes you write Python, understand evaluation metrics and have already produced a model, however modest. What it adds is the engineering layer around it: experiment tracking with MLflow, the move from notebook to parameterised script, sweep jobs for hyperparameters, pipelines broken into reusable components, then full automation with GitHub Actions and GitHub environments that gate promotions to production.
The shift to generative AI happens at module 9. The last six modules treat Microsoft Foundry agents with the same rigour: agent specifications, prompt versioning in GitHub repositories, structured evaluation experiments with scoring rubrics, batch evaluations in Python wired into continuous integration, monitoring of key metrics, and tracing to debug complex call chains. That continuity between classic machine learning and generative AI is what makes the programme distinctive.
In French speaking Switzerland, AI projects are leaving the proof of concept stage. Banks, insurers, medtech, watchmaking and public bodies share one constraint: a model that influences a decision must be traceable, explainable and reproducible, with a clear record of what was deployed, when and by whom. Expectations around data governance and the documentation of automated systems keep rising, and they translate directly into what this course teaches: a model registry, versioned pipelines, archived evaluations and continuous monitoring.
That is also why local teams now look less for people who can train a model than for people who can keep it in service. The Microsoft Certified: Machine Learning Operations Engineer Associate certification, prepared by the AI-300 exam, targets exactly that role.
AI-300 pairs well with two neighbouring courses. AI-103, Develop AI applications and agents on Azure, focuses on building: Azure AI services, agent orchestration, retrieval augmented generation. It answers what you build, while AI-300 answers how you run it. Taking AI-103 first is the natural order, though the reverse works well for someone coming from a DevOps background.
DP-750, Implement data engineering solutions using Azure Databricks, covers the upstream side: ingestion, transformation and dataset reliability. An MLOps pipeline is never better than the data feeding it, and many production incidents blamed on the model actually originate in the data chain. Together the two paths cover the full journey for a team.
The AI-300 modules are freely readable. Time is not. Setting up an Azure Machine Learning workspace yourself, connecting GitHub Actions to Azure with a federated identity, running a sweep job and then a pipeline, and finally deploying a monitored endpoint takes weeks of trial and error. In four guided days, with the official Microsoft labs and a certified trainer, you run the whole chain and meet the usual breaking points: missing permissions on a resource, a badly registered artefact, a metric absent from the dashboard, a promotion blocked by a GitHub environment.
The other benefit is judgement. Should retraining be triggered by a drift threshold or by a schedule? How many examples belong in an agent evaluation set? When is an automated evaluator reliable enough to block a release? Those questions get settled by discussing your own cases with the trainer and the other participants, not by reading a documentation page.
The most profitable habit is to take, within the following week, a model already in service at your organisation and apply the method to it: register it, wrap it in a parameterised training script, wire a first GitHub Actions workflow, then add monitoring. One complete loop on a simple case is worth more than a target architecture drawn on paper. For agents, start by moving prompts out of application code and versioning them, since that single step unlocks everything else.
If you are aiming for certification, plan to sit the AI-300 exam within a few weeks of the session, while the tooling is still fresh.
Do I need to know how to train a model before attending?
Yes. The course assumes you are comfortable with Python and have already produced a model, even a simple one. AI-300 deals with everything that happens after training.
Does the course cover generative AI or only classic machine learning?
Both. The first eight modules address MLOps with Azure Machine Learning, the following six cover GenAIOps with Microsoft Foundry: prompts, evaluations, monitoring and tracing of agents.
Do I need prior GitHub Actions experience?
Basic Git knowledge is enough. The GitHub Actions workflows are built step by step during the labs, including the secure connection to Azure and environment based promotion gates.
How is this different from AI-103?
AI-103 teaches you to build AI applications and agents on Azure. AI-300 teaches you to deploy, evaluate and monitor them. The two courses follow on from each other naturally.
Is the certification mandatory?
No. Sitting the AI-300 exam is optional and is scheduled separately from the training session.
Can I attend remotely?
Yes. The course is offered on site in Geneva and Lausanne, and as a virtual class with the same trainer and the same labs.
Monday to Friday
8:30 AM to 6:00 PM
Tel. 058 307 73 00
ITTA
Route des jeunes 35
1227 Carouge, Suisse
Monday to Friday, from 8:30 am to 06:00 pm.
