AI-200 cheat sheet: Azure AI Cloud Developer Associate

Develop containerized solutions on Azure

Azure Container Apps: a serverless container platform built for microservices and event-driven apps, with built-in autoscaling (including scale-to-zero) — a common host for lightweight AI inference or retrieval services.
Azure Kubernetes Service (AKS): full Kubernetes control for more complex, multi-service AI workloads that need custom orchestration beyond what Container Apps offers out of the box.
Azure Container Registry (ACR): private storage for container images, with vulnerability scanning and geo-replication for multi-region deployments.
Dockerfile and image-build basics still apply — a developer should be able to containerize an application, push it to ACR, and deploy it to either target.

Vector search and embeddings: converting content into vector representations and storing them for similarity search — Cosmos DB's vector search capability and Azure AI Search both support this pattern.
Retrieval-augmented generation (RAG): retrieve relevant context from a vector store, then pass it to a language model alongside the user's prompt — grounds the model's output in your own data rather than relying purely on what it learned during training.
Azure Managed Redis: a managed caching layer that can also serve as a fast vector store for smaller-scale or latency-sensitive retrieval workloads.
Data partitioning and indexing choices affect both retrieval accuracy and cost — know the trade-off between broader recall and tighter, faster queries.

Azure AI Foundry (and the underlying Azure OpenAI / Azure AI services) expose models and AI capabilities through REST APIs and SDKs — know how a client application authenticates to and calls these services.
API Management can front AI service endpoints for rate limiting, key rotation, and usage tracking across multiple consuming applications.
Event-driven integration: Event Grid and Service Bus for decoupling an AI pipeline's stages (ingest → process → notify) rather than chaining everything synchronously.
SDKs vs. direct REST calls: SDKs handle auth, retries, and serialization for you — prefer them unless there's a specific reason to call the REST API directly.

Managed identities let an Azure resource authenticate to another Azure service (like an AI endpoint or a Key Vault secret) without embedding credentials in code or config.
Azure Key Vault stores API keys, connection strings, and secrets centrally, with access controlled through Azure RBAC rather than shared out-of-band.
Application Insights and Azure Monitor provide request tracing and performance telemetry — essential for diagnosing latency in a multi-step AI pipeline (is the slowdown in retrieval, the model call, or the response formatting?).
Rate limiting and quota management matter specifically for AI service calls, which are often metered per-token or per-request and can fail distinctly from a typical API timeout.

Reading a summary is the easy part. Take the free 10-question diagnostic to see which of these domains you actually have, then drill the rest.