How microsoft/call-center-ai works
Project type and purpose
• This is a cloud-native, serverless “Call Center AI” solution written in Python.
• It exposes a POST /call API to initiate AI-driven voice calls (outbound/inbound), streams voice in real time, leverages Azure Communication Services for telephony, Azure Cognitive Services for speech and translation, Azure OpenAI for LLM, and persists call transcripts, claims, reminders, and syntheses in Cosmos DB (with Redis caching, Storage queues, Event Grid).
• It also hosts a simple public website (Jinja2 templates) to display call reports.
File structure
• Top-level:
– cicd/ (Dockerfile, Bicep IaC templates, versioning)
– app/ (Python application)
· helpers/ (config, logging, LLM calls, event handling, monitoring)
· models/ (Pydantic domain models for call, claim, message, reminders)
· persistence/ (interfaces and adapters for Cosmos DB, Redis, Azure Queue Storage, Twilio/SMS)
· resources/ (static assets and templates for public website, local token files)
· main.py (FastAPI or similar entrypoint)
– public/ (static assets)
– config-*.yaml (local and remote config examples)
– docs/, examples/, tests/
– GitHub Actions workflows, .devcontainer, Makefile, pyproject.toml
README for additional insights
• Technologies: Azure Container Apps, Communication Services, Cognitive Services (STT/TTS/Translation), OpenAI GPT-4.1/4.1-nano, Cosmos DB, Azure Storage Queues, Event Grid, Redis, Application Insights.
• It uses Bicep for IaC and GitHub Actions for CI/CD, with a single Docker image deployed to Azure.