Email Automation — B-Yond Inbox Classifier

Polls a Microsoft 365 inbox via the Graph API and automatically classifies inbound form-submission emails using an LLM. Results are logged to an Excel file.

Microsoft 365 Inbox
        │
        ▼
  email-poller          polls on a configurable interval
        │
        ▼
email-classification-agent   LangGraph graph hosted on langgraph-api
        │
        ├── extracts JSON payload from the email body
        ├── researches the company online
        └── returns structured classification (qualify/disqualify, company type, countries, etc.)
        │
        ▼
  Excel tracker         appends each result to output/email_classifications.xlsx

The poller also triggers the email-ingestion-agent graph for emails that contain attachments (PDFs, PowerPoints), which are extracted and passed to the LLM alongside the email body.

• Docker & Docker Compose
• A Microsoft 365 account with an App Registration in Azure AD (for Graph API access)
• An OpenRouter API key (or OpenAI-compatible endpoint)

In Azure Active Directory, register an app and grant it the following application (not delegated) permissions:

• Mail.Read
• Mail.ReadBasic

Note down the Client ID, Client Secret, and Tenant ID.

Create a .env file in the project root:

# Microsoft Graph / MSAL
CLIENT_ID=
CLIENT_SECRET=
TENANT_ID=
MAIL_USER=inbox@example.com          # mailbox to poll

# LLM (OpenRouter or OpenAI-compatible)
OPENROUTER_API_KEY=
LLM_MODEL=google/gemini-2.5-flash   # or any OpenRouter model

# LangGraph Postgres
POSTGRES_PASSWORD=

# LangSmith (optional tracing)
LANGSMITH_API_KEY=
LANGSMITH_PROJECT=email-automation

A full list of supported config keys is in src/agent/config.py.

docker compose up --build

On first run Docker will build the two images (Dockerfile.agent and Dockerfile.email-poller).

Key polling settings in .env (or environment):

Classified emails are appended to output/email_classifications.xlsx with columns including:

• Thread ID, Email ID, Sender, Subject
• Action (qualify / disqualify)
• Company Name, Company Type, Operation Countries
• Contact Name, Contact Email, Salesperson
• Confidence score, Date of Contact

Edit the prompt files in src/agent/prompts/:

• email_classification_system_prompt.txt — system instructions for the LLM
• email_classification_user_prompt.txt — user message template (supports {email_subject}, {email_body}, {attachment_text})

src/agent/get_access_token.py is a standalone script for testing Graph API authentication:

python src/agent/get_access_token.py

Prints a short preview of the acquired token (or NO TOKEN on failure).

src/agent/
├── email_poller.py               # Entry point — polls inbox, dispatches to graphs
├── email_classification_graph.py # LangGraph classification graph
├── email_ingestion_graph.py      # LangGraph ingestion graph (attachments)
├── graph_schemas.py              # Pydantic models / LangGraph state
├── config.py                     # All configuration (pydantic-settings)
├── excel_tracker.py              # Appends results to Excel
├── logger.py                     # Centralised logging
├── get_access_token.py           # Standalone token testing utility
└── prompts/                      # LLM prompt templates