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 project: a video compliance auditor

BrandGuardian is the project this course ends on: it takes the URL of a brand or influencer video, reads what the video says and shows, and checks it against a set of advertising rules, returning every violation it finds. The video calls it an Azure multimodal compliance orchestration engine, built with LangGraph and LangSmith.

Last updated: 15 Oct, 2026 · LLMOps

Swapping the stand-ins for the real thing ended the hands-on thread on a keyless shop assistant you could run. This part is the other half: a full production project, operated the way the rest of the course describes, shown end to end. It runs on Azure and needs an account, so it is walked through rather than run here, the same as the deployment lessons.

The project and its architecture · from the End to End Multimodal LLMOps Project video · 6:00 to 9:00

What it audits, and why

The rules come from two PDFs the video uses as the guiding documents: FTC guidance on influencer advertising, the kind that says a paid promotion has to be disclosed. The job is to take a video ad and decide whether it follows those rules: is the disclosure there, are the claims it makes allowed. A human reviewer doing this reads the rules, watches the ad, and writes up what is missing. BrandGuardian does the same three steps in software.

The shape of the system

  • A request comes in at a FastAPI endpoint with a video URL. The video skips a Streamlit UI by design, because a production backend is the part that matters.
  • The video is processed into something a model can read: Azure Video Indexer pulls its transcript, the text on screen through OCR, and its metadata.
  • The rules are retrieved from a knowledge base: the FTC PDFs, embedded and stored in Azure AI Search, queried for the passages that bear on this ad.
  • An auditor model, GPT-4o on Azure, compares the ad against the retrieved rules and returns the violations it finds, each with a category, a severity and a description.
  • Everything is watched: LangSmith traces the chain of thought, and Azure Application Insights records every request, its latency and its errors.
The BrandGuardian architecture: a POST /audit request with a video URL fans into an index path, where Azure Video Indexer produces transcript and OCR and Blob Storage holds the video, and a rules path, where FTC rule PDFs are embedded into Azure AI Search. Both feed the auditor node, GPT-4o comparing the ad against the retrieved rules, which emits violations by category and severity. A dashed band marks the index and auditor nodes as the LangGraph workflow, and an observability strip underneath notes LangSmith traces each audit while Azure Application Insights records every request.

Why it is the course's capstone

  • It is every lifecycle stage at once: a data pipeline, a model choice, a built workflow, evaluation of the output, deployment behind an API, and observability over the lot.
  • It is multimodal: the input is a video, not text, so the data stage does real work before the model ever sees it.
  • It is operated, not demoed: it has a knowledge base that is versioned, traces you can open, a latency you can measure and errors that are recorded.
Watch out. This project runs on paid Azure services and needs an account, keys and roles set up. It is shown here, grounded in the video, rather than run, the same as The Azure deployment from the project. Read it for the shape of a real LLMOps system; the keyless shop assistant earlier in the course is the one you run yourself.
Try it yourself
  • Pick a short influencer video you have seen and list, in one line each, the rules an auditor would check it against.
  • Sketch the five stages above as boxes and mark which one turns the video into text.
PreviousSwapping the stand-ins for the real thing

Little by little, you're building something great.