Anna Veronika Dorogush is founder and CEO of Recraft, an AI lab building image models and tools for visual creators. Previously she led the machine-learning tools team at Yandex and helped create CatBoost, an open-source gradient-boosting library for working with structured data and categorical features.[1]
Start with CatBoost: gradient boosting with categorical features support →
Career
Dorogush graduated from Moscow State University’s Faculty of Computational Mathematics and Cybernetics and the School of Data Analysis. Her work included ABBYY, Microsoft and Google before Yandex, where she helped develop CatBoost and led the machine-learning tools team. At Recraft, she leads a lab whose models power both a design application and an API.[1][4][5]
Contributions
Handle categorical data inside gradient boosting
With Vasily Ershov and Andrey Gulin, Dorogush authored a paper explaining CatBoost’s implementation and categorical-feature support. It describes processing categorical values during training and provides CPU and GPU implementations. The work connects algorithm design to a usable library, with evaluations across several public datasets rather than one application-specific model.[3]
Reduce leakage while learning from training examples
She also coauthored the NeurIPS CatBoost paper with Liudmila Prokhorenkova, Gleb Gusev, Aleksandr Vorobev and Andrey Gulin. The paper analyzes prediction shift caused by target leakage and develops permutation-based ordered boosting and categorical-feature processing. To reduce leakage, an example’s categorical statistics use preceding examples in a permutation, excluding its own target. The paper supplies theoretical analysis and experiments, explaining the library’s approach.[2]
Ideas
Recraft’s V4.1 announcement distinguishes expressive generation, more predictable utility output and native vector generation. Dorogush describes aesthetics as a goal alongside prompt accuracy. The application and API let creators use those models in visual work; the models and tools are the work of the Recraft team, rather than a solo research contribution.[5]
Tools, writing & talks
Writing, research & talks
Start here
A practical account of CatBoost’s categorical-feature handling and CPU/GPU implementations.
Start with the explanation of categorical features and the library’s design before the benchmark tables. Familiarity with decision trees and supervised learning helps; continue to the NeurIPS paper for the more detailed treatment of leakage and ordered boosting.
First author with Vasily Ershov and Andrey Gulin[3]- CatBoost: unbiased boosting with categorical featuresproceedings.neurips.cc
The theoretical and experimental account of ordered boosting and categorical target statistics.
Coauthor[2]