Case study

Inbox Sentinel: Email Classification and Routing

An IMAP monitor that reads every inquiry landing in RJ Autonomous' mailboxes, classifies it with Claude, and routes it to the right person around the clock. We built it for ourselves and run it every day.

Internal system, running 24/7Automation · Internal tooling
StackPython, IMAP, Anthropic Claude
StatusInbox Sentinel is the automation we use on our own business.
Quick overview

Internal system, running 24/7

Inbox Sentinel is the automation we use on our own business. It is also the reference implementation for the intake and routing work we build for clients.

The problem

What needed to be fixed.

A shared inbox is where inquiries go to wait. A real lead sits next to a vendor pitch and a password reset, and whoever checks first has to read all of it. Response time depends on who is at their desk.

What we did

How we tackled it.

A Python service watches the mailboxes over IMAP. Each new message is classified by Claude (new inquiry, existing client, vendor, billing, noise) and routed: the right person is notified with a summary and a suggested next step, and the original stays untouched in the mailbox.

How we built it

The approach and structure we used.

Here's how we thought about the implementation, the choices we made, and how we delivered it, without sharing anything that would compromise client privacy.

01

Connected to the mailboxes over IMAP with read-only access, so the system can never delete or send mail on its own.

02

Defined a small set of categories and a routing rule for each, written down before any model was involved.

03

Used Claude for classification and summary, with the category list and examples in the prompt so results stay consistent.

04

Logged every decision so misroutes can be reviewed and the categories tightened over time.

The stack

Tools and platforms we used.

PythonIMAPAnthropic Claude
The results

What this created for the business.

  • Every inquiry is read and routed within minutes, at any hour, with no ticket queue to check.
  • Leads reach the person who handles them with a summary already written.
  • The same pattern powers the intake automation we build for service firms.
About the metrics

We share what matters, keeping the sensitive details private.

This is our own system, so the volumes are small and the point is the pattern: classify with a fixed category list, route by rule, keep a human in charge of the reply.

Ready when you are.

No pitch decks. No twelve-step sales funnel. A conversation about what you need and whether we're the right fit.