I spent a few months fine-tuning smaller models for domain-specific tasks instead of relying on large general-purpose APIs. Here is what actually worked, what did not, and when it is worth the effort.