Models have become highly capable despite noise in their training data. That noise now sets a ceiling on their performance. The next level of model performance will be unlocked by training data with zero noise.

Expert credentials and existing QA practices are probabilistic proxies for data quality. Decca does not rely upon imperfect proxies. It does the work required to guarantee the quality of the training data it delivers.

Decca stands apart among training-data companies because it was founded from within the work. As a legal engineer, Afzal Hasan has written and reviewed legal training data and led thousands of lawyers attempting to make frontier models fail. It was through that work that he identified the untapped opportunity the company is built around. His career spans private practice in securities law, general counsel, and serial entrepreneurship, including building a unicorn.

Decca is in Toronto.