Fighting Email Spam on Your Mail Server with LLMs — Privately

unixfox

I’m sharing a blog post I wrote: https://cybercarnet.eu/posts/email-spam-llm/

It’s about how to use local LLMs on your own mail server to identify and fight email spam.

This uses Mailcow, Rspamd, Ollama and a custom proxy in python.

Give your opinion, what you think about the post. If this could be useful for those of you that self-host Mailcow.

Thanks

ETNyx

Hi man,

thanks for sharing your work.

Could you provide us with some comprehensive statistics from those six months? The screenshot you shared shows 15 rejections. From those, only 5 seem to have been reclassified by the GPT plugin by default, if I understand correctly. That would mean around 66% are irrelevant, and if those are sent to the GPT plugin, that’s wasted money.

Secondly, regarding the financial side, could you elaborate on your costs and setup? At first glance, if you’re not using any spare GPUs in your office or homelab, renting them would cost around 10–50 times more than my current setup. That wouldn’t be feasible for me. Do you have any price estimates for using a SaaS (API) version instead?

unixfox

ETNyx Hello,

I shared on reddit:

In one month I got 165 spam email rejected by the classic rspamd rules and 35 rejected by the AI analyzing it. Out of 935 emails received.

https://old.reddit.com/r/LocalLLaMA/comments/1o3xluf/fighting_email_spam_on_your_mail_server_with_llms/niyosj7/

Regarding the financial side, I’m lucky to be able to use a free tier of https://fly.io which provide dedicated GPU where I installed Ollama (billed per seconds). But this shouldn’t cost that much for personal usage if you are using a SaaS provider (inference API).
I just tried the latest prompt from a previous email scan by an AI on https://openrouter.ai, and it costed me $0.0000532 with the model Gemma 3 12B.
So in my case, if I have around 40 emails scanned by AI per month, it would cost me $0,002128/month or $0,025536/year.