What does Palantir's whiteboard architecture actually consist of?

There are three distinct layers. The first is studying language and heuristics, which is the early Palantir approach to understanding unstructured data. The second is an ontology layer where data gets combined with algorithms and deterministic solutions. The third layer is where they now incorporate LLMs. Palantir calls the combination of these three layers sovereign AI. It is different from just calling the OpenAI API or using Claude directly because at no point in that stack does your raw proprietary data leave your environment. You are using the model as a tool rather than feeding the model as a training source.