CertifyDeckAI-300 has five connected domains spanning machine learning operations and generative AI operations. The official outline effective March 5, 2026 gives the greatest weight to model lifecycle operations while also testing Foundry, evaluation, observability, RAG, and fine-tuning.
1. Design and implement MLOps infrastructure (15–20%)
Workspaces, datastores, compute, identity, data assets, environments, components, registries, Bicep, Azure CLI, and GitHub Actions all appear.
Practice secure GitHub access, private networking, source control, and decisions about sharing, versioning, isolation, and workload identities.
2. Implement model lifecycle and operations (25–30%)
The largest domain covers MLflow, AutoML, notebooks, hyperparameter tuning, scripts, distributed training, pipelines, job comparison, registration, feature retrieval specifications, Responsible AI, and archiving.
Production topics include online and batch endpoints, managed inference, troubleshooting, progressive delivery, rollback, drift, monitoring, alerts, and retraining triggers.
3. Design and implement GenAIOps infrastructure (20–25%)
Configure Foundry resources and projects, managed identities, RBAC, private networks, Bicep, and Azure CLI.
Also cover foundation model selection and deployment, serverless versus managed compute, model versions, provisioned throughput, prompt variants, comparison, and Git-based prompt management.
4. Generative AI quality and observability (10–15%)
Build evaluation datasets and mappings; distinguish groundedness, relevance, coherence, and fluency; evaluate harmful content; and automate built-in or custom metrics.
For operations, know continuous monitoring, latency, throughput, response time, token and resource cost, logs, traces, and debugging.
5. Optimize generative AI systems (10–15%)
Tune similarity thresholds, chunks, retrieval strategies, embeddings, and hybrid search. Validate improvements with relevance metrics and A/B tests.
Understand advanced fine-tuning, synthetic data, performance monitoring, and the customized model lifecycle from development through production.
Turn the outline into a schedule
Weight time by domain but complete at least one hands-on exercise per objective. Follow the AI-300 study plan and track objective, lab, and evidence.
Treat the Microsoft study guide as authoritative and check its date before booking.
Frequently asked questions
How many AI-300 domains are there?
Five, covering MLOps infrastructure, ML lifecycle, GenAIOps infrastructure, generative quality and observability, and generative optimization.
Which AI-300 domain is largest?
Machine learning model lifecycle and operations at 25–30%.
Is Microsoft Foundry included?
Yes. Foundry is central to GenAIOps infrastructure, model deployment, prompt management, evaluation, and monitoring.