Building MAE: My Agentic Employee

December 10, 2025

Motivation

We all know the struggle. You sign up for one service, and suddenly you’re on five mailing lists. You pay for groceries, and the receipt sits in your inbox forever. I was overwhelmed by the sheer volume of daily emails - newsletters, receipts, and promotions burying the actual important messages. I got tired of manually filtering my inbox, flicking repeatedly to delete or move most of them.

While cloud-based solutions exist, I wanted a system that:

  1. Runs locally without sending data to third-party AI services.
  2. Is extensible and customizable.
  3. Runs on cheap, easily replaceable hardware.
  4. Costs negligible amounts to operate (vs $240/year for SaaS).
  5. Keeps me independent of AI cloud providers.

Enter MAE (My Agentic Employee)

MAE is a lightweight, intelligent email processing agent designed to run on edge devices. I use a Radxa Zero 3 (RK3566), a capable yet affordable single-board computer.

It uses a local NPU-accelerated AI model to classify emails and perform automated actions, keeping my inbox clean without sending my data to third-party AI services.

Capabilities

  1. Filters the Noise: Automatically moves promotions to trash.
  2. Organizes Clutter: Archives receipts and newsletters (Feeds) so they’re accessible but unobtrusive.
  3. Highlights Importance: Leaves personal and work emails in the Inbox.
  4. Respects Privacy: Runs entirely locally.

Intelligent Classification

The agent fetches emails via the Gmail IMAP API and categorizes them into four buckets:

  • Transactions: Bills, invoices, payments, stock trades.
  • Feed: Newsletters, updates, announcements.
  • Promotions: Marketing emails, offers etc.
  • Inbox: Personal correspondence, urgent work.

Automated Actions

Once classified, MAE takes action:

  • Promotions -> Moved to Trash (recoverable for 30 days).
  • Transactions & Feed -> Marked as Read and Archived.
  • Inbox -> Left untouched.

This simple logic drastically reduces the cognitive load of checking email.

MAE Solution Architecture

  • Hardware: Radxa Zero 3 (RK3566)
  • AI Inference: RKNN-based 1 TOPS NPU
  • Model: MobileBERT (58 MB)
  • Training: PyTorch -> ONNX -> RKNN
  • Inference Time: ~700 ms/email
  • CPU Load: ~1% average (7% peak)
  • RAM Usage: ~100 MB average (200 MB peak)

The heart of MAE is a MobileBERT model optimized for the Rockchip NPU (Neural Processing Unit) using the RKNN Toolkit. This allows the agent to classify emails with high accuracy and incredibly low power consumption, running efficiently on a small single-board computer.

Two rounds of training the model with labelled data gives an accuracy of around 98.6%. I have not had the need to move any deleted email out of the trash bin, and I’ve had to actually read through only 3-4 emails per day after taking this live.

Future Development

  • Calendar Integration: Extract dates/times and create Google Calendar events.
  • Historical Processing: Clean up years of neglected inbox clutter.
  • Contact Management: Extract and save contact info to Google Contacts.
  • WhatsApp Integration: Create a “network map” from WhatsApp messages.