How Does Deel Train Its Compliance Model?

Aaron Goldsmid says Deel trains small language models on a wiki built from the company's own compliance knowledge. On global work and compliance, that small model is at least as good as Sonnet and is quickly approaching Opus.

Deel has thousands of compliance experts across its markets. The knowledge base grew organically, then the team put it into a standard wiki based on Wikipedia so people could reach it easily and treat it as a source of truth. Language models are focused on the corpus, and small language models are now trained on the same knowledge base. The upside is second-level responses versus waiting five to 10 to 15 or 20 seconds, at a fraction of the cost.

The investment in the wiki is what builds a moat. The in-house model performs well and stays highly reliable because it is trained on a known good set of data rather than the whole universe. A general model can be asked to look only in one place, but that limit cannot be enforced.