Right now, it’s a blend: structured pre-processing + context engineering + proprietary prompt frameworks on top of Gemini + Web Search, with fallback to OpenAI and Claude for certain synthesis tasks. Over time, we may shift more of that into a vertically-trained model to improve speed, privacy, and cost.