AI-300 Tests Shipping and Running Models Safely — Not Building Them from Scratch
AI-300 is the MLOps and GenAIOps exam that followed DP-100. Think like the engineer who owns workspaces, endpoints, Foundry deployments and evaluation gates, not the data scientist tuning one notebook.
Check Your Readiness →Most candidates understand Operationalizing Machine Learning and Generative AI Solutions (AI-300) concepts — and still fail. This exam tests how you apply knowledge under pressure.
Scenarios set a constraint such as no stored secrets, EU-only processing, zero-downtime release or rising token spend, then ask for the feature or sequence that meets it with least risk and effort. 700/1000 to pass in 120 minutes; Microsoft does not publish the question count.
Split 10% of live traffic to green
Mirror a percentage of traffic to green: mirrored responses are discarded, so callers only ever get blue's answers. When ready, shift the traffic split gradually and keep blue until the rollout is proven, since it is the fast rollback path.
Deploy to a managed online endpoint and loop over rows
Use a batch endpoint, which scores asynchronously on compute clusters that scale to zero between jobs. Tune mini_batch_size and instance count for throughput. Online endpoints keep instances running and suit low-latency, per-request scoring.
Store a service principal ID and secret as repository secrets
Add an OIDC federated credential to a Microsoft Entra identity with the subject scoped to the main branch or a protected environment. GitHub exchanges its token for an Entra token, so there is nothing to store or rotate.
Use a global standard pay-per-token deployment
Choose a data zone provisioned throughput deployment in the EU. PTUs give reserved capacity, predictable latency and spend for steady load; overflow returns 429 unless spillover is set. Bursty, low traffic would fit standard deployments better.
Switch to a larger embedding model
Exact identifiers match best on keyword BM25 retrieval, so enable hybrid search merged by reciprocal rank fusion, optionally with the semantic ranker. The ranker only reorders results already retrieved, and a new embedding model forces re-embedding the corpus.
Azure Machine Learning sweeps support grid, random and Bayesian sampling, but early termination (Bandit, median stopping, truncation) cannot be combined with Bayesian. To cut tuning cost with early stopping, switch to random sampling.
Identity-based datastores check the data plane, so the user or compute identity also needs Storage Blob Data Reader or Contributor on the storage account. AzureML Data Scientist on the workspace alone still fails to read the data.
Global deployments may process requests in any Azure region. A data zone deployment keeps processing inside a geography such as the EU or US, and a regional deployment pins it to one region. Residency wording rules global out.
Groundedness scores whether claims are supported by the retrieved context and needs the context column mapped. Relevance only checks the answer addresses the query; coherence and fluency judge structure and language. AI-assisted evaluators also consume judge-model tokens.
Facts that change every few weeks belong in a retrieval index, which also supports citations. Supervised fine-tuning shapes format, tone and behaviour; DPO needs preferred-versus-rejected pairs. Fine-tuned deployments are billed hourly while deployed, used or not.
Strengthen weak areas with exam-style practice questions and detailed explanations.
Simulate the real exam experience and assess your readiness under timed conditions.
Review key concepts, objectives, and exam topics in one place.
Get personalized explanations, learning recommendations, and instant answers.
Follow a structured learning path designed to help you prepare efficiently.
AI-300 rewards production-grade MLOps and GenAIOps choices. See which deployment calls you would get wrong.