Overview
Built for my own workflow on personal infrastructure. It polls Office 365 every 15 minutes, classifies each email with Claude, and either creates an Asana task or archives it based on configurable routing rules.
The Problem
The daily inbox was a time sink:
- Critical requests buried in promotional noise
- Action items forgotten in long thread chains
- Context lost when forwarding emails to project tools
- Hours wasted on manual sorting and task creation
The Solution
A system that runs continuously on personal VPN infrastructure, processing emails every 15 minutes:
Intelligent Classification
Claude AI analyzes each email against a configured skill matrix, determining:
- Category: Client request, internal task, notification, spam
- Priority: Urgent, high, normal, low
- Routing: Which team member or workflow should handle it
- Action Type: Requires response, task creation, or archive
Automated Task Creation
When an email requires action, the system:
- Extracts key details (deadline, stakeholders, requirements)
- Links relevant SOPs for self-service resolution
- Creates properly formatted Asana tasks
- Attaches the original email for context
- Assigns based on skill-based routing matrix
Technical Architecture
Key Components
Email Monitor Service
- Systemd service running on Debian 12 VPN server
- Polls Microsoft Graph API every 15 minutes
- Tracks processed emails to prevent duplicates
AI Classification Engine
- Claude 3.5 Haiku for cost-effective classification
- Custom prompt engineering for business context
- Structured output parsing for reliable routing
Rules Database
- SQLite for routing rules and client associations
- Configurable skill matrix for team assignments
- SOP linking for common request types
Results
The system has been running in production since late 2024:
- ~30 minutes saved daily on email processing
- 15-minute polling cadence for new messages
- Actionable messages routed into Asana with the original email attached
- Scheduled operation with email failure alerts
Technical Highlights
- Stateless Design: Each poll is independent, enabling easy recovery
- Failure Resilience: AWS SES notifications on any processing errors
- Context Preservation: Full email thread linked to every task
- Cost Optimized: Haiku model keeps AI costs under $5/month