Azure Multi-modal
Compliance Ingestion Engine
using LangGraph
AI-powered video compliance auditing at scale
Automated. Traceable. Intelligent.
- Project Overview
The Azure Multi-modal Compliance Ingestion Engine is a sophisticated AI system designed to automatically audit video content for regulatory compliance. Instead of having humans watch hours of video footage, this system uses artificial intelligence to analyze videos, extract information, and detect compliance violations in seconds.
The system automates compliance auditing by ingesting video files from YouTube, CCTV, or local storage, extracting all useful information (spoken words, on-screen text, detected objects), analysing this against compliance rules using semantic AI, and generating timestamped reports with severity ratings and corrective actions.
- System Architecture Diagram
The diagram below shows the complete architecture of the Azure Multi-modal Compliance Ingestion Engine. It illustrates how the four major zones — Entry Points, Orchestration, Azure Infrastructure, and External Intelligence — are interconnected.
https://www.youtube.com/watch?v=rV3HJ4LEZ7k&t=170s
- System Architecture Components
The architecture is divided into five functional zones. Each zone has a distinct responsibility in the compliance pipeline:
3.1Entry Points
Two modes of interaction are supported:
3.2Orchestration Layer
The orchestration layer contains four tightly coupled components that form the main processing pipeline, as shown in the architecture diagram:
RAG Workflow / LangGraph — manages the stateful AI agent loop with decision-making and retries
Video Processor (yt-dlp + Indexer) — downloads and submits video to Azure Video Indexer, extracts insights
Retrieval Engine — converts queries to vectors, searches Azure AI Search, returns relevant chunks
Compliance Auditor — receives evidence + rules, reasons with GPT-4o, outputs structured violations
3.3Azure Infrastructure
3.4External Intelligence + Observability
- Orchestration — LangGraph & RAG
4.1LangGraph — Stateful AI Workflow
LangGraph enables the compliance auditor to behave as an adaptive agent rather than a linear script. The agent can loop back when evidence is insufficient:
4.2RAG Architecture
RAG (Retrieval-Augmented Generation) is the core pattern enabling the AI to reason over video content it was not trained on:
- Complete End-to-End Processing Flow
The following 9 stages describe how a raw video is transformed into a structured compliance report: