Hey @thunderbird
If you want, you can try out #alectors, my NLRL library, for #spam or #fraud #detection on #emails for #thundermail.
Check it out at https://apotheke.earth/docs/alectors
It uses a pretrained #encoder to embed natural language into vector space, and then passes the vectors through a #transformer-based PPO agent, so that the output can have a custom number of dimentions. This means that rather than having an #LLM agent with a #prompt "[...] PLEASE OUTPUT PROPER JSON [...]", you can ensure that the output will be one of three, four, or as many actions.as you define.
The only downside is that you will have to train the agent yourself, as we as come up with an environment (I am working with a friend/colleague on custom nlrl env api, but its pre-prealpha).
Concerning the license, it is #sourcefirst, meaning that it has a clause for #fair #compensation regardimg #commercial #use. For this usecase, by #thunderbird, I would be perfectly happy with simply a #thundermail account so I can leave #fastmail and still use all two of my domains for email.