Azure AI Foundry: Enterprise Model Infrastructure & Fine-Tuning Blueprint¶
Real Production Infrastructure HITRUST CSF v11 Aligned Simulated Clinical Datasets HIPAA Zero-PHI Guardrails ARB Approved
1. Enterprise AI Foundry Architecture & Topology¶
The Azure AI Foundry Enterprise Model Infrastructure establishes a sovereign, zero-trust foundation for hosting, fine-tuning, and evaluating Large Language Models (LLMs) and multi-modal agents in regulated healthcare environments.
Dedicated Standalone AI Model Factory Repository
The full codebase, OpenTofu infrastructure modules, Python SDK fine-tuning runners, and CI/CD policy gates for this architecture are maintained in the dedicated repository:
👉 FreeFades2Black/mosaic-azure-ai-model-factory
Infrastructure Integrity & Data Notice
Real Azure Architecture: The Terraform module, Azure AI Foundry Hub, OpenAI cognitive deployments, Azure AI Search service, and Private Link endpoints described here represent 100% real, deployable cloud infrastructure.
Simulated Data: Clinical training datasets, protocol questions, and evaluation scenarios are simulated / synthetic data strictly designed to avoid exposure of real Protected Health Information (PHI).
graph TD
subgraph "Healthcare Edge & Consumer Apps"
ClinicianPortal["Clinical Decision Support Web App<br/>(Private Endpoint / TLS 1.3)"]
EHRStream["EHR Interoperability Pipeline<br/>(HL7 v2 / FHIR R4 Event Hubs)"]
end
subgraph "Azure AI Foundry Enterprise Hub & Project (East US 2)"
AIHub["Azure AI Foundry Hub<br/>(Top-Level Governance & RBAC)"]
AIProject["Clinical Agent Project<br/>(mosaic-clinical-agent-project)"]
AIHub --> AIProject
subgraph "Foundation Model Deployments"
GPT4o["GPT-4o Multi-Modal Model<br/>(GlobalStandard 100K TPM)"]
EmbeddingModel["text-embedding-3-large<br/>(3072 Dimensions)"]
end
subgraph "RAG & Safety Guardrails"
AISearch["Azure AI Search<br/>(Hybrid Vector Search + Semantic Ranker)"]
ContentSafety["Azure AI Content Safety<br/>(Prompt Shield + Harm Filters)"]
end
subgraph "Supporting Infrastructure"
Storage["Geo-Redundant Storage (ADLS Gen2)<br/>(Datasets, Prompty & Checkpoints)"]
KeyVault["Gunslinger Key Vault<br/>(FIPS 140-2 Level 3 HSM CMK)"]
AppInsights["Application Insights & LAW<br/>(730-Day Audit Log Stream)"]
end
end
ClinicianPortal --> ContentSafety
ContentSafety -->|Pass| GPT4o
AISearch -->|Vector Grounding| GPT4o
EmbeddingModel --> AISearch
GPT4o --> Storage
GPT4o --> KeyVault
GPT4o --> AppInsights2. Core Pathways for Building Models in Azure¶
When building and customizing models in Microsoft Azure, three primary patterns are supported:
Pathway A: Fine-Tuning Foundation Models (Azure AI Foundry SFT / LoRA)¶
For domain adaptation where clinical terminology or EHR format adherence is required:
1. Dataset Preparation: Curate synthetic training pairs in JSONL format conforming to the Chat Completion format (system, user, assistant).
2. Job Submission: Dispatch fine-tuning jobs via the Azure AI Python SDK or CLI with hyperparameter constraints (epochs, batch size, learning rate multiplier).
3. Model Checkpointing & Evaluation: Checkpoint fine-tuned weights directly into encrypted Azure Blob Storage with automated Groundedness and Content Safety evaluations.
Pathway B: Retrieval-Augmented Generation (RAG) with Azure AI Search¶
For dynamic clinical knowledge retrieval without updating foundational model weights:
1. Document Ingestion: Chunk and vectorize clinical policy manuals and HITRUST controls using text-embedding-3-large.
2. Vector Indexing: Store 3072-dimensional embeddings in Azure AI Search using the HNSW cosine similarity algorithm.
3. Semantic Reranking: Apply Microsoft's deep learning Semantic Ranker to reorder top-$k$ retrieved chunks before prompt injection.
Pathway C: Custom ML Model Training (Azure Machine Learning)¶
For traditional tabular classification, regression, and deep learning (PyTorch / Scikit-Learn):
1. Compute Clusters: Provision auto-scaling GPU compute clusters (Standard_NC6s_v3 or Standard_ND96asr_v4).
2. MLflow Tracking: Stream metrics, loss curves, and confusion matrices to Azure ML Workspace.
3. Managed Online Endpoints: Deploy models behind blue/green deployment traffic routers.
3. Reference Terraform Blueprint (terraform/azure-ai-foundry-enterprise/)¶
This architecture is packaged as an automated Terraform / OpenTofu module:
# Azure AI Foundry Enterprise Hub Instance
resource "azurerm_ai_foundry" "hub" {
name = "${var.prefix}-aihub-${var.environment}"
location = var.location
resource_group_name = var.resource_group_name
storage_account_id = azurerm_storage_account.ai_storage.id
key_vault_id = var.key_vault_id
application_insights_id = azurerm_application_insights.ai_insights.id
public_network_access = "Disabled"
identity {
type = "SystemAssigned"
}
}
# Clinical AI Project
resource "azurerm_ai_foundry_project" "clinical_project" {
name = "mosaic-clinical-agent-project"
location = var.location
ai_services_hub_id = azurerm_ai_foundry.hub.id
identity {
type = "SystemAssigned"
}
}
4. Python SDK Orchestration (src/ai_foundry/agent_fine_tuning_and_rag.py)¶
You can execute model fine-tuning and RAG indexing programmatically:
# Execute the AI Foundry Fine-Tuning and RAG Vector Indexing Pipeline
python src/ai_foundry/agent_fine_tuning_and_rag.py
5. Sandboxed Code Interpreter Tool & Clinical Agent¶
To eliminate LLM arithmetic hallucinations when dealing with critical dosage equations, laboratory indices, and statistical modeling, the model architecture integrates a sandboxed Python Code Interpreter Tool:
- Deterministic Execution: Offloads complex clinical mathematics to an isolated Python AST runtime rather than relying on generative token probabilities.
- Built-in Clinical Formula Library: Includes audited implementations for:
- eGFR (Estimated Glomerular Filtration Rate): Cockcroft-Gault equation with female gender coefficient (0.85).
- BSA (Body Surface Area): Mosteller formula ($\sqrt{\frac{h \times w}{3600}}$) for oncological dosage calculations.
- MME (Morphine Milligram Equivalents): Multi-factor opioid conversion matrix.
- Anion Gap: Serum electrolyte balance equation ($[\text{Na}^+] - ([\text{Cl}^-] + [\text{HCO}_3^-])$).
- AST Security Guardrails: Strict Abstract Syntax Tree analysis blocks dangerous imports (
os,sys,subprocess,socket), disk I/O, and eval/exec operations.