AZURE AI/ML ECOSYSTEM — CRASH COURSE | Novant Health Interview Prep May 2026 | 4Shadow Analytix — Confidential

Mental Model — What Azure AI Is

Azure AI is Microsoft's answer to the question: what if every enterprise AI capability — model serving, MLOps, RAG, data pipelines, governance, and healthcare integration — ran inside the same identity and billing envelope as Office 365, Teams, and Epic's DAX Copilot? The strategic bet is that healthcare orgs already run Microsoft, so the AI layer should be native, not bolted on. At Novant, this is already true: DAX Copilot is a Microsoft product. Truveta runs on Azure. Every AI vendor Novant has deployed runs on Azure infrastructure. The job is to architect into that ecosystem, not alongside it.

The Two Core Azure AI Control Planes

Platform	What It Does	AWS Equivalent
Azure AI Foundry	GenAI/LLM serving, RAG, agents, model catalog (1,800+ models), prompt engineering, evaluations	Bedrock + SageMaker JumpStart combined
Azure Machine Learning	MLOps — training pipelines, model registry, managed endpoints, experiment tracking (MLflow native), responsible AI dashboard	SageMaker Pipelines + Experiments + Model Registry + Endpoints

Key difference: AWS separates GenAI (Bedrock) from MLOps (SageMaker). Azure merges them under Foundry/Azure ML with a shared control plane.

AWS → Azure Translation Map (Your Stack → Their Language)

You Built This (AWS/RunPod)	Say This at Novant (Azure)
vLLM on RunPod H100	Azure ML Managed Online Endpoints / Foundry model deployment
MLflow on PostgreSQL backend	Azure ML Experiments — MLflow is native first-class citizen
Polars bronze/silver/gold pipeline + psycopg3 COPY	Azure Data Factory pipelines → ADLS Gen2 (Data Lake Storage)
Anthropic Claude API (4-agent eval pipeline)	Azure OpenAI Service (GPT-4.1 / o3 / o4-mini)
Alation semantic layer (J&J, Evernorth)	Microsoft Purview — catalog, lineage, PHI classification, access policy
Cloudera → AWS S3/Redshift/Databricks (J&J)	Azure Databricks (same engine, first-party native) → Azure Synapse / Fabric DW
AWS S3 storage layer	Azure Data Lake Storage Gen2 (ADLS Gen2) / OneLake (Fabric)
Bedrock Knowledge Bases + OpenSearch	Azure AI Foundry RAG + Azure AI Search (hybrid BM25 + vector + semantic reranker)
IAM roles	Microsoft Entra ID + Azure RBAC
CloudTrail + CloudWatch	Azure Monitor + Azure Activity Log
GitHub Actions CI/CD	GitHub Actions (Microsoft owns GitHub — even more native on Azure)

Key Concepts — 8 Terms Every Architect Must Own

1. Azure AI Foundry

The unified GenAI control plane: model catalog (1,800+ models), managed online endpoints, prompt flow, RAG pipelines, agent orchestration (Foundry Agent Service — GA), evaluations, and fine-tuning. Think of it as Bedrock + SageMaker Studio + LangChain runtime in one Azure-native product. At Novant: this is where LLM-based clinical documentation, CDI AI, and future agentic workflows would be architected.

Note: GPT-4o retired March 2026. Current flagship: o3/o4-mini (reasoning), GPT-4.1 (generation).

2. Azure ML (MLOps Layer)

Persistent compute clusters, YAML-first pipeline definitions, model registry with staging/production versions, managed online endpoints (blue/green deployments), batch endpoints, and a Responsible AI Dashboard combining fairness, error analysis, counterfactual what-if, and feature importance in one view. Critical difference from SageMaker: compute clusters persist — build auto-shutdown policies or costs continue running.

3. Microsoft Fabric

A SaaS lakehouse that bundles what AWS requires 6–8 services to assemble: OneLake (unified data lake, like S3 but cross-workspace), Data Factory (ETL), Spark notebooks, T-SQL Data Warehouse (serverless, on Delta Lake), Real-Time Intelligence (streaming, Kusto/KQL engine), Power BI (native BI), and Purview integration. One SKU. Delta Lake as universal format. At Novant: CitiusTech is likely building the enterprise data layer on Fabric/Synapse.

4. Microsoft Purview

Enterprise data governance: automated scanning across Azure + AWS S3 + on-prem SQL, 200+ built-in PHI/PII classifiers (critical for HIPAA), business glossary with stewardship workflows, end-to-end lineage from ADF pipelines through Synapse to Power BI, and access policy enforcement. Your equivalent experience: Alation at J&J and Evernorth. Purview is what Alation would look like if it were native to the entire Azure data stack.

5. Azure AI Search

The RAG retrieval layer: four-layer hybrid search — BM25 keyword → vector → reciprocal rank fusion → Azure OpenAI semantic reranker. Key differentiator: Entra ID permission-aware retrieval — a user querying a SharePoint-indexed knowledge base only sees documents their Entra ID entitles them to. No custom Lambda. In healthcare, where record-level access control is mandatory, this matters architecturally.

6. Azure Health Data Services

Managed FHIR R4/R4B server with SMART on FHIR native (AWS HealthLake requires you to build the auth server yourself), DICOM service for medical imaging, and MedTech service for IoT device → FHIR mapping. At Novant: Epic runs on Virtustream but data flows via FHIR APIs into Azure — this is the integration layer. DAX Copilot is a Microsoft product running on this infrastructure.

Novant's Actual Azure Stack — What You'd Be Working With

Layer	Tool	What It Does at Novant
EHR	Epic on Virtustream	640+ locations; Virtustream is managed hosting — not Azure, but FHIR APIs flow out to Azure
Clinical AI	DAX Copilot (Microsoft/Nuance)	Ambient documentation in Epic; 900 clinicians, 550K+ encounters; expanding to ED + inpatient
Cloud	Azure (primary)	All major AI vendor workloads (Jvion, Truveta, DAX) run on Azure
Federated Query	Starburst (Trino)	"Point users to data where it lives" — federates Azure + on-prem + Epic Caboodle/Clarity; this is NOT a centralized warehouse yet
Enterprise Data Layer	CitiusTech (SI, Sept 2024)	External preferred vendor building analytics layer; architect role is the internal counterpart
Research Data	Truveta on Azure	Novant is founding co-owner; 130M+ patient RWD; Truveta Intelligence launched April 28, 2026
Revenue Cycle AI	SmarterDx	CDI, 19 hospitals, 100% chart review, doubled ROI forecast
Governance	Not yet named	Purview is the natural fit given Azure dominance — likely next governance layer

Starburst insight: The architectural principle Justin Byrd (VP Data Platform) stated explicitly — "point users to data where it lives." Do NOT frame a centralized lake as the solution. Frame federation-first, then governed access layers on top.

Watch-Outs — What NOT to Say

Don't lead with Snowflake or Databricks as assumed tools — no confirmation Novant uses either. Their architecture is Starburst + Azure + Epic Caboodle.
Don't call it "SageMaker" — translate in your head before speaking: managed online endpoints, Azure ML pipelines, Foundry deployments.
Don't imply you'd build a centralized lake — their VP said explicitly they federate. Propose federation-first architecture, then governed access layers.
Don't position AWS as primary — your J&J and Evernorth work used AWS and Databricks, which is fine to cite. Frame the patterns as cloud-portable, then name the Azure equivalent.
GPT-4o is retired (March 2026). Say o3, o4-mini, or GPT-4.1 — not GPT-4o.

Validation / Testing Angle — Your QA Background Applied

The Responsible AI Dashboard in Azure ML is the governance artifact a QA-trained architect owns — it combines fairness analysis, error analysis, counterfactual what-if, and feature importance in a single auditable view. This is what GxP validation looks like in Azure's language.

Azure AI Content Safety + Prompt Shields = the automated hallucination and jailbreak gate. At NewsRx, you built this with HHEM-2.1-Open + BERTScore. Azure packages the equivalent as a managed deployment-level filter.
Azure ML experiment tracking (MLflow native) = what you built with MLflow on PostgreSQL. Same tracking paradigm — runs, params, metrics, artifacts, model registry — just managed inside Azure workspace.
Purview lineage = the audit trail across ADF pipelines through Synapse to Power BI. At J&J you built this in Alation. Purview automates the crawling and lineage capture you did manually.
SMART on FHIR in Azure Health Data Services = the access control gate at the clinical data boundary. Entra ID enforces who sees which patient records — the same Human-in-the-Loop governance principle you applied at Invistics and Amgen, now enforced at the data layer.

Interview Language — Say These Phrases

"My current production stack runs on the AWS and RunPod side — MLflow for experiment tracking, vLLM for inference, Polars medallion pipelines into PostgreSQL. Those patterns are cloud-portable. At Novant, that translates to Azure ML for MLOps, Azure AI Foundry for LLM serving, Azure Data Factory for pipeline orchestration, and Purview for the governance layer. Same architecture, Azure-native tooling."

"The Starburst architecture Novant runs — federating Azure, on-prem, and Epic Caboodle without moving data — aligns with how I'd approach this. You don't centralize first and govern later. You govern access at the federation layer and build the analytics surface on top."

"Microsoft Fabric is the direction — OneLake as the unified storage layer, Data Factory for ingestion, Synapse or Fabric DW for analytics, Power BI native for reporting, Purview baked in for governance. That's what I'd architect the CitiusTech data layer to grow into."

"Responsible AI in Azure isn't a dashboard bolted on after deployment. The RAI Dashboard in Azure ML gives you fairness, error analysis, counterfactual what-if, and feature importance in one auditable view at model registration time — before the endpoint goes live. That's the same design principle I used at Amgen and Invistics: governance at architecture time, not review time."

"DAX Copilot expanding from 900 clinicians to ED and inpatient is an architecture problem before it's an adoption problem. The integration layer between Epic, Azure OpenAI Service, and the clinical workflow is where the friction lives — and that's where I'd focus the first 90 days."