Video Compliance AI
Video Compliance AI is an agentic system that automatically audits YouTube videos for FTC regulatory compliance. Give it a YouTube URL and it returns a structured audit report detailing any violations found — missing disclosures like #ad or #sponsored, misleading claims, and unsubstantiated endorsements. No human needs to watch the video.
What It Does
Brand teams and agencies manually review influencer videos for regulatory compliance — slow, inconsistent, and expensive. This system automates the full pipeline: download the video, extract speech and on-screen text, retrieve relevant FTC rules, reason over the evidence, and return a structured JSON audit report.
How It Works
Azure Video Indexer downloads the video via yt-dlp and simultaneously performs speech-to-text transcription and OCR on on-screen text. This multimodal ingestion captures both spoken claims and text overlays. The transcript and OCR output are combined into a query that retrieves the top 3 relevant FTC rule chunks from Azure AI Search, which holds the FTC Disclosures guide and YouTube Ad Specifications as indexed vectors. GPT-4 then reasons over the video content against those retrieved rules and returns a structured compliance verdict.
The entire pipeline is orchestrated by a two-node LangGraph DAG — an indexer node handles ingestion and an auditor node handles RAG reasoning. Both nodes share a TypedDict state object.
What Makes It Interesting
The system is deployed as a production FastAPI service with Azure Monitor and OpenTelemetry auto-instrumentation capturing request latency, error rates, and dependency traces. LangSmith provides per-node agent trace inspection for debugging. Authentication uses DefaultAzureCredential — in development it uses Azure CLI, in production it switches automatically to Managed Identity with no code change.
Tech Stack
Python, LangGraph, LangSmith, Azure Video Indexer, Azure AI Search, Azure OpenAI (GPT-4), Azure Monitor, OpenTelemetry, FastAPI, Pydantic, yt-dlp.