Maxim Kan

CV Engineer

Wildberries

Wildberries is a major Russian online marketplace, connecting shoppers with a broad catalog of goods and sellers across multiple countries.

Built and deployed product image defect detection, virtual try-on, and image text translation systems, with a 15% metric gain and 95% text accuracy.

• Defect detection: automated visual quality control of the product. • Designed and trained a CV model to detect visual defects in product images across large-scale catalog inventory. • Built the training pipeline, labeling workflow, and evaluation protocol for continuous model improvement. • Deployed the model into the moderation pipeline, improved the baseline model metrics by 15%. • Virtual try-on of clothes: generative preview of garments on customers. • Developed virtual try-on system enabling customers to preview clothing items before purchase.

CV Engineer

PRISMA

Built personalized image generation and synthetic data pipelines, and researched diffusion inversion methods for image reconstruction and editing.

• Personalized generation: subject-driven personalized image generation. • Implemented personalized DreamBooth fine-tuning to generate images of a customer from a few reference photos. • Optimized the fine-tuning procedure for training time and output fidelity to support a consumer-facing feature. • Synthetic dataset generation: generative data augmentation for downstream CV models. • Built generative pipelines to synthesize training data for custom effects. • Balanced and augmented datasets used by downstream computer-vision models.

ML Engineer

Phygitalism

Phygital+ is a cloud service for creating visual content with neural networks, including GANs, super-resolution, diffusion models, and style transfer.

Built cloud tools for running and controlling generative models, trained models for custom domains, and designed a DeepFake pipeline for an award-winning project.

• Phygital+: cloud service for creating visual content via neural networks. • Developed a system for cloud-based launch of neural networks such as GAN, Super Resolution, Diffusion models, Style Transfer, etc. • Adapted algorithms and added custom parameters for better user control of generated results. • Trained generative models for custom domains. • VHUTEMAS: DeepFake to replace the lecturer with a famous artist. • Designed the DeepFake pipeline. • Red Dot Award-winning project.

Researcher

ADASE Group

Developed a CycleGAN-based image translation method for unsupervised enhancement and super-resolution of depth sensor maps.

• Unsupervised enhancement and super-resolution of depth maps obtained with a Structure depth sensor. • Developed a CycleGAN-based image-to-image translation scheme and adapted it to the depth domain.

Master of Science, Mathematics and Computer Science

Skolkovo Institute of Science and Technology (Skoltech)

Bachelor's degree, Electronics and Nanoelectronics

Tomsk State University of Control Systems and Radioelectronics