Azure Services Deep-Dive Q&A (+ AI-102 cram)

The JD names Azure services as "THE MUST." Expect service-by-service "do you actually use this?" probing — specific features, gotchas, current names. This file is the rapid drill, plus an AI-102 skills cram. If an interviewer asks "have you used [service]," you want a fluent, specific answer, not a generic one.

Table of Contents

• 1. Azure OpenAI Service
• 2. Azure AI Search
• 3. Azure AI Document Intelligence
• 4. Azure AI Vision
• 5. Azure AI Foundry / Agent Service
• 6. Azure Cosmos DB
• 7. Azure Container & platform services
• 8. Supporting AI services
• 9. The renamed-services table (don't get caught)
• 10. AI-102 skills cram

1. Azure OpenAI Service

Use it fluently: resource → deployment (model + version + SKU); call the deployment name. SKUs: Standard (PAYG), Global/Data-Zone Standard, Provisioned (PTU). Quotas: TPM/RPM; handle 429 with backoff. Auth: Managed Identity (Cognitive Services OpenAI User), keyless. Features: chat completions, function/tool calling, structured outputs (JSON Schema), vision (GPT-4o), embeddings, content filters (configurable), on-your-data (built-in AI Search RAG), prompt caching, batch API. Gotchas: pin api-version and model version; "deployment name ≠ model name"; abuse-monitoring opt-out for sensitive data; not all models in all regions. Likely Q: "How do you deploy and call GPT-4o?" → resource in a residency-compliant region → create gpt-4o deployment (PTU for prod) → call by deployment name with Managed Identity, pinned API version, Content Safety on, traced to App Insights.

2. Azure AI Search

Use it fluently: index (fields: searchable/filterable/vector), vector profile (HNSW, cosine), semantic configuration. Query: hybrid (BM25+vector, RRF) + semantic ranker (cross-encoder) + filters (security trimming) + captions/answers. Ingestion: indexer + skillset + integrated vectorization (chunk + embed for you), or push API. Newer: agentic retrieval (LLM query planning over the index). Gotchas: it's AI Search not "Cognitive Search"; semantic ranker is a paid tier; replicas (query throughput/HA) vs partitions (index size); filter syntax for ACLs (search.in). Likely Q: "vector or hybrid?" → hybrid+semantic, and explain RRF + retrieve-then-rerank.

3. Azure AI Document Intelligence

Use it fluently: Read (OCR), Layout (prebuilt-layout — tables/structure, Markdown output), prebuilt (invoice/receipt/ID/bank-statement/pay-stub), custom (template vs neural), composed (auto-route), classifier. Output: fields + confidence + bounding regions; async poller. Studio for labeling/human review. Gotchas: it's Document Intelligence not "Form Recognizer"; use Layout for RAG chunking; confidence-gate high-risk fields to human review; Arabic support varies by model/version. Likely Q: "build a KYC pipeline" → classify → prebuilt/custom extract → confidence gate + human review + validation (IBAN mod-97) + audit.

4. Azure AI Vision

Use it fluently: Read OCR (print/handwriting, many languages incl. Arabic), Image Analysis 4.0 (Florence: captions/tags/objects/smart-crop/OCR), Custom Vision (train a classifier/detector with few samples), Face (gated), Video Retrieval. Gotchas: Vision Read vs Document Intelligence OCR (use Document Intelligence for documents); Face is limited-access/compliance-sensitive. Likely Q: "OpenCV vs Vision vs GPT-4o?" → OpenCV preprocess, Vision/Doc-Intelligence OCR, GPT-4o for judgment.

5. Azure AI Foundry / Agent Service

Use it fluently: unified portal/SDK; model catalog (Azure OpenAI + open/partner models as serverless or managed compute); evaluators (groundedness/relevance/safety, built-in + custom; offline + continuous/online); prompt flow; Foundry Agent Service — managed agent runtime on the OpenAI Responses API, with built-in tools (file search, code interpreter, AI Search, MCP, functions), memory (procedural/user/session), multi-agent orchestration, observability to Azure Monitor; governance/identity; publish to Teams/M365. Gotchas: "Azure AI Foundry" rebranded "Microsoft Foundry" (2026); SK/AutoGen → Microsoft Agent Framework. Likely Q: "Foundry Agent Service vs self-hosted LangGraph?" → managed memory/observability/governance vs full control/portability; interop via MCP.

6. Azure Cosmos DB

Use it fluently: NoSQL (Core) API, JSON docs in containers; partition key (high cardinality, even access, single-partition queries, ≤20GB/partition); RU/s (provisioned/autoscale/serverless), 429 on overage; consistency (Session default; Strong↔Eventual spectrum); TTL; native vector search (DiskANN, VectorDistance()); change feed; global distribution. Gotchas: partition key is effectively irreversible; cross-partition queries cost; "Document DB" in the JD = this. Likely Q: "partition key for chat?" → conversationId for single-partition point reads; justify.

7. Azure Container & platform services

Use it fluently: Container Apps (ACA) (serverless, KEDA autoscale, revisions/traffic-split, VNet, Dapr) — default for the API; AKS (full k8s, GPU pools) when needed; ACI (simple jobs); ACR (registry + scan); APIM (auth/throttle/audit); Front Door (global LB/WAF). Identity: Entra, Managed Identity, RBAC, Key Vault, Private Endpoints. Delivery: GitHub Actions/Azure DevOps, Bicep, App Insights/Monitor, Defender/Sentinel. Likely Q: "ACA vs AKS?" → ACA by default (serverless/autoscale/managed); AKS for existing k8s/GPU/complex topology.

8. Supporting AI services

• Azure AI Language — PII detection/redaction, NER, sentiment, key-phrase, language detection, CLU (intent). Use for PII redaction + Arabic/English detection.
• Azure AI Translator — Arabic↔English when you must pivot to one language.
• Azure AI Speech — STT/TTS for IVR/voice banking (incl. Arabic).
• Azure AI Content Safety — harm categories (0–7), Prompt Shields, Groundedness detection, protected material.
• Microsoft Presidio (OSS) — customizable PII redaction for banking entities.

9. The renamed-services table (don't get caught)

Using a current name unprompted signals you live on the platform; using an old one signals you've been away.

10. AI-102 skills cram

⚠️ AI-102 retires 2026-06-30 (16 days out). Don't promise to sit it; reference the skills, and check Microsoft Learn for the Foundry-era successor. The exam's five domains map almost exactly onto this role — a good last-minute checklist:

1. Plan & manage an Azure AI solution (25–30%) — choose services; provision (multi-service vs single); security (Managed Identity, Key Vault, Private Endpoints, RBAC); responsible AI/content safety; monitor & manage cost; CI/CD. → K00, K07, K09.
2. Generative AI solutions (now emphasized: GenAI + agentic) — Azure OpenAI deployments, prompt engineering, RAG / on-your-data, agents (Foundry Agent Service), evaluation, content filters. → K01, K02, K03, K06, K08.
3. Natural language processing (25–30%) — Azure AI Language: NER, PII, sentiment, key-phrase, language detection, CLU, translation, speech. → K05 §6, K08 §8.
4. Azure AI Vision (15–20%) — Image Analysis, Read OCR, Custom Vision, Face; Document Intelligence (layout/prebuilt/custom). → K04, K05.
5. Decision support / knowledge mining (10–15%) — AI Search (indexers, skillsets, vector/semantic), Content Safety. → K02.

If you know this track, you know the AI-102 surface — and more (the agentic/banking depth the exam only touches).