[systemz]

Telegram seems like strange choice but ok. I personally started to tinker with something similar - IMAP client which learns how you categorize emails into folders and helps with that. I started with manual rules and want to connect some optional automagic later.

I noticed that some emails are valid for only some hours and then it's in inbox unnecessarily giving that +1 to unread badge.

On other hand server side filtering works before arriving in inbox so you don't see for example, that parcel will arrive today. Next day you need to manually archive it. After 10-20 mails like that, if you are busy - half of inbox needlessly needs user action if you are into inbox zero.

[est]

> Telegram seems like strange choice but ok

if you think about it, Telegram seems like the one of the few mobile chat apps allowing easy bot integration.

[lxgr]

Why should this be a mobile chat app integration at all, though?

[est]

what alternatives do you have in mind?

ChatGPT's success comes with a chat based UI, every LLM AI agent these days had a hidden chat context somewhere. So choosing a chat app is pretty intuitive and straightforward.

[olivierl13]

Well that's a valid point, I might add other communication channel later, maybe the community will add a new one ;)

Interesting point - for the moment all actions are manually validated on this project but I might also add features to make the agent autonomous to some extent.

[bgro]

I haven’t used AI for sorting email yet but have over many attempts over many years tried to setup manual filters.

I try to make them as absolutely bulletproof as possible but I’m still running into problems because email is so rough.

Critical Microsoft emails for example constantly have random sketchy looking domains but are legit.

Or someone named “Don Otre Plyson” gets caught in the “DoNotReply” filter and I never see his important message.

I don’t know if I’d blindly trust AI to guess patterns better than me where I’d have no way of ever knowing if I potentially missed something.

[systemz]

Yeah, I don't trust AI magic either. Very precise manual filters can run out of date quickly, "AI" can be unpredictable. Maybe naïve bayes will be a middle ground?

I use and like SpamSieve for macOS but it's limited to spam/not spam. I would really like support for more categories than two.