Impressive work. I’ve experimented with similar setups using LLaMA models locally and found them useful but still limited for nuanced financial reasoning. Curious how you fine-tuned your model for domain-specific language — was it rule-based, RAG, or some mix?