– Trained Hibou: a family of Foundational Vision Transformers for digital pathology using a modified DINOv2 framework on a proprietary 1.2 billion image dataset. Achieved the strongest metrics on both tile-level and slide-level public pathology benchmarks for Hibou-L.– Authored a paper detailing training procedures and metrics. Contributed to open-source by releasing the Hibou-B model on GitHub and Hugging Face under the Apache 2.0 license. Hibou-B is the strongest freely available ViT-Base model in digital pathology.– Leading the development of multiple SOTA AI-driven algorithms for cancer classification and segmentation on gigapixel histopathological slides. Responsible for idea generation, research, integrating open-source solutions, developing custom DL architectures, training models, and successfully deploying them to production.

Detection/classification/segmentation. Inference on devices: distillation, pruning, quantization.

Shadow Fight: Arena– Added quantization to production version on devices, increased generalization of neural networks, rebuilt distillation pipeline from the scratch. All of that allowed to use less neural networks in production,which reduced space complexity from O(N^2) to O(N), where N is the number of characters in the game (Hundreds MBs of saved space) – Achieved a fine-grained behavior control of RL bots through reward engineering.AI Video MoCap for Cascadeur– Developed a 3D motion capture from a monocular video– Researched SOTA papers, tailored public models to our requirements, trained private models to further improve the quality of results– Prepared the solution for production.

– Published a paper in a Q1 journal, the first author.– Solved a problem of cross-sensor calibration between different satellites. Developed training and inference pipelines from scratch. Optimized inference to work on gigapixel satellite imagery that doesn’t fit onto a single GPU.