It's not a large project, but you would need fine-tuning or a lot of few-shot examples to approach SpamAssassin's efficacy. It would still be slower than SpamAssassin. This is one of those "X but with LLM" ideas :)

As far as I know, Gmail has a neural net which is trained by examples from people reporting spam. That is a more potent approach. If you're going to restrict the output grammar to 'true or false', why even run a giant net? You could achieve the same loss with a much smaller one.