Event-to-worker orchestration
Graph recording notifications, online-meeting and recording API retrieval, containerized services, and event-triggered Azure Container Apps Jobs.
CASE STUDY / 01AZURE / MICROSOFT GRAPH / AI
From conversation to intelligence.
AI Meeting Intelligence Platform
An event-driven pipeline that turns Teams meeting recordings into the foundation for transcription, summaries, and actionable meeting intelligence.
PROJECT STATUSEngineering Project / Proof of Concept
The challenge.
A meeting recording is only the start. Retrieving the right artifacts, preparing media, coordinating asynchronous processing, and recovering from failure all need to work before an AI workflow can produce useful meeting intelligence.
THE OBJECTIVE
Build a webhook-driven Teams recording pipeline that separates event intake, media processing, and downstream intelligence, with identity and operational controls designed into each stage.
Follow the system.
Select a stage to explore its role in the architecture.
SYSTEM FLOW / SELECT A STAGE01 OF 06
STAGE 01
Microsoft Graph
Recording notifications trigger intake; subscription renewal keeps the event connection active.
STAGE 02
Azure webhook
A containerized .NET 10 / C# service accepts recording events and hands work to a queue.
STAGE 03
Storage Queue
Asynchronous dispatch decouples event delivery from media processing and isolates failures.
STAGE 04
Container Apps Job
An event-triggered worker retrieves meeting metadata and recordings, then prepares and chunks audio with FFmpeg.
STAGE 05
Blob Storage
Prepared artifacts are stored for downstream processing using managed identity authentication.
STAGE 06
Meeting intelligence
Power Automate and speech workflows connect prepared audio to transcription, summaries, and Minutes of Meeting workflows.
What I built.
Media preparation
Recording metadata processing, FFmpeg conversion, audio chunking, and Blob Storage handoff, with multilingual processing considerations.
Operational foundations
Subscription renewal automation, Log Analytics, Azure Monitor, operational alerts, and poison-message handling.
The choices that matter.
Decouple intake from processing
Queue-based work distribution isolates expensive media tasks from the webhook and gives failures a defined recovery boundary.
Design for repeated delivery
Idempotency and safe deployment considerations are part of the processing design; an event should not imply uncontrolled repeated work.
Keep authentication secretless
Managed identities replace embedded Azure credentials across the service and storage boundaries.
Controls, by design.
- Managed identities and scoped access to Azure resources.
- Scoped access to meeting artifacts and storage through application permissions and managed identities.
- Failure isolation and poison-message handling make unsuccessful processing visible.
A foundation for what’s next.
An engineering proof of concept connecting recording events to media preparation and downstream meeting-intelligence workflows. I built the cloud, identity, and observability foundations and explored the transcription and AI handoffs.
Architecture is generalized. Client and operational details remain private.