Most AI training code is just a big for loop with each line calling highly performant c/cpp libraries underneath. There is no value that go or rust can add here.

Regular llama might already work well. You can generate query with schema. Openai api is much cheaper if you don’t have much scale.