I kept missing messages I actually cared about.
Not because I was ignoring them. Because by the time I had space to reply, too much time had passed and it felt weird. The thread died. The relationship quietly suffered.
I did not need a better inbox. I needed another me.

The Actual Problem
Most AI writing tools generate something that sounds fine but reads like nobody. The goal with Mae was different: if the person receiving the reply could not tell I did not write it, it worked. If it felt slightly off, it failed.
That bar is harder than it sounds.

What I Built
Mae connects to Gmail, WhatsApp, and Slack. When a message comes in it reads the sender, pulls the relationship history, and generates a reply in my voice.
The voice part required training on real sent messages, segmented by relationship type. How I write to a close friend is different from how I write to a cold contact. The model learns both and switches automatically.
The thing that actually made it useful was confidence scoring. Mae rates its own drafts before sending. High confidence goes out automatically. Low confidence comes to me. Users tune the threshold themselves.
That shift from reviewing everything to trusting the right things is what changed it from a demo to a tool.

Mae is live at runmae.ai across Gmail, WhatsApp, slack and more.

ai #machinelearning #buildinpublic #productivity