Fair point. I used SOTA models for the analysis to prioritize quality, but since the heavy media processing is local, API costs stay negligible (or free). The architecture is modular, though—you can definitely swap in a local LLM for a fully air-gapped setup.