Yeah LLM seems like the go to solution. And the best one. And talking about resources, we can use barely smart models which can generate coherent sentences, be it 0.5b-3b models offloaded to CPU inference only.