start with something more simple. Just use multiple choice in the background like an old choose your own adventure book but the AI makes the choices and pair that with a similar character complexity and scope. Use a multi model architecture where a small manager limits length and constrains to theme.

Models are already developing an internal character profile for everyone in a prompt context. All you need to do is add some meta questions about the probable education level of the textual responses to determine age within reason. The model picks up on word choice and grammar well. There is a major component of dogma involved in all interactions where the model is trying to infer probable expectations. In fact, if you avoid describing characters and infer all within verbal dialog, it unlocks a lot more dynamic range of behavior.

I do not think 3rd party junk is the way to go. Code it to own and utilize the local model ecosystem. Pitch it to sell bleeding edge hardware as a leading AAA. You will never get a consistent long term fine tune behavior from any API. You can fine tune train a local model to behave as expected 95+ percent of the time, which is better than any general API model.