LLMOpsThe LLMOps lifecycle · MLflow 3 · FastAPI · Docker · AWS/Azure
Dashboard
0%
1
Curious builder0 XP earned · 300 to level 2
0 daysFinish a lesson to begin
Badge collection0 of 6 unlocked
50 small wins to finish your pathNext lesson →

The Azure services the project runs on

BrandGuardian is assembled from managed Azure services, one per job: Blob Storage for the video, Video Indexer to read it, AI Search for the rules, Foundry for the model and the embeddings, and Application Insights to watch it all. Each is provisioned once and reached through a connection string in the environment.

Last updated: 15 Oct, 2026 · LLMOps

The project: a video compliance auditor drew the architecture. This lesson names the services behind each box, because the first real work of the project is provisioning them and putting their credentials in one place.

The services and the project dependencies · from the End to End Multimodal LLMOps Project video · 44:00 to 47:00

One service per job

ServiceIts job in the project
Azure Blob StorageStores the unstructured data: the downloaded video itself.
Azure Video IndexerReads the video: transcript, on-screen text by OCR, and insights.
Azure AI SearchThe knowledge base: the vector store the FTC rules are embedded into.
Azure Foundry / Azure OpenAIThe model: GPT-4o as the auditor, and text-embedding-3-small for the rules.
Azure Application InsightsThe monitor: latency, errors and user activity across every request.
LangSmithThe trace: the chain of thought for one audit, request by request.
The six Azure services BrandGuardian provisions, in a grid, each with its job: Azure Blob Storage stores the downloaded video, Azure Video Indexer gives transcript OCR and insights, Azure AI Search is the rules knowledge base, Azure Foundry serves GPT-4o and text-embedding-3-small, Application Insights tracks latency errors and user activity, and LangSmith traces one audit. A strip below notes each is reached through a connection string or key in the environment and needs an access role assigned by hand in the portal.

The model, from Foundry

The video deploys two models in Azure Foundry: GPT-4o, which it calls the brain of the system, as the auditor, and text-embedding-3-small to turn the rule documents into vectors. Both are named deployments; the app reads the endpoint, the key and the deployment name from the environment rather than hard-coding them, the same pattern as Config, not hard-coded constants.

bash
# the model, from Foundry
AZURE_OPENAI_ENDPOINT=...
AZURE_OPENAI_API_KEY=...
AZURE_OPENAI_CHAT_DEPLOYMENT=gpt-4o
AZURE_OPENAI_EMBEDDING_DEPLOYMENT=text-embedding-3-small

# the knowledge base and the storage
AZURE_SEARCH_ENDPOINT=...
AZURE_SEARCH_API_KEY=...
AZURE_SEARCH_INDEX_NAME=brand-compliance-rules
AZURE_BLOB_CONNECTION_STRING=...

# reading the video, and watching the app
AZURE_VIDEO_INDEXER_ACCOUNT_ID=...
APPLICATIONINSIGHTS_CONNECTION_STRING=...
LANGCHAIN_API_KEY=...
LANGCHAIN_PROJECT=brand-guardian
Watch out. Every one of these is a billed resource that also needs an access role assigned, which the video sets up by hand in the Azure portal: a Video Indexer contributor role, a storage role, and so on. Provisioning is most of the project's first hours, and tearing the resources down afterwards is the same discipline the SageMaker endpoint needed in Deploying an open-source LLM on SageMaker.
Try it yourself
  • Match each of the six services to the lifecycle stage it serves, from data to monitoring.
  • Name which two services would keep billing if you forgot to tear them down.
PreviousThe project: a video compliance auditor

You understood something today that you didn't yesterday.