We are a team striving to disrupt the key processes of the Bank and its partners through the implementation of breakthrough computer vision solutions. Our approach combines three key elements:
- Cutting-edge models — we use state-of-the-art architectures, including multimodal solutions and LLMs, not afraid to experiment with the latest research
- Powerful harness — we create our own infrastructure for rapid prototyping, training, and deployment of models, which accelerates the development cycle many times over
- Deep expertise — we rely on accumulated experience in implementing CV solutions in production, from prototype to scaling on edge devices.
Our projects cover a wide range of areas: from image and video analysis to building autonomous systems and World Models.
Responsibilities
- analyzing requirements and designing solution architecture at the intersection of classic CV and Vision LLM
- developing hybrid computer vision models: from traditional tasks (detection, segmentation, classification, tracking) to multimodal VLLM solutions for scene understanding and visual reasoning
- integrating VLLM into CV pipelines for Visual Question Answering, Image / Video Caption, cross-modal search, and complex image analysis tasks
- optimizing models for industrial implementation: quantization, distillation, inference acceleration for operation on edge devices;
- interacting with the development team to transfer models into production, including API development and integration with existing systems
- documenting experiments, metrics, and results (MLflow, Tensorboard, WandB)
- code review, mentoring junior specialists, and maintaining code quality within the team
- researching and implementing new approaches from scientific publications on CV and VLLM.
Requirements
- experience in Data Science / Computer Vision — from 3 years, with a portfolio of implemented projects
- strong engineering skills, understanding of IT development processes and DevOps processes
- deep knowledge of Python and core ML/CV libraries: PyTorch, OpenCV, Transformers
- understanding of classic computer vision fundamentals: filters, morphology, keypoints, feature matching, epipolar geometry;
- experience with Vision LLMs and multimodal architectures (CLIP, BLIP, LLaVA, Flamingo, etc.)
- knowledge of core neural network architectures for CV tasks (CNN, Vision Transformer, DETR, YOLO, SegFormer) and understanding of their integration principles with language models
- experience with model deployment frameworks (ONNX, TensorRT, OpenVINO, TorchScript) and inference optimization
- ability to read scientific articles in English (CVPR, ICCV, ECCV, NeurIPS) and quickly prototype new approaches
- understanding of Large Language Model principles: fine-tuning, prompt engineering, LoRA, adapters
- experience with vector databases and cross-modal search will be an advantage.
Will be a plus:
- publications or presentations at industry conferences
- experience with video analytics and object tracking
- Kubernetes skills
- experience with geospatial data or satellite imagery
- knowledge of MLOps fundamentals and experience building pipelines for model training and deployment.
Conditions
- comfortable modern office near Kutuzovskaya metro station
- hybrid work format
- annual salary review. Annual bonus
- corporate gym and recreation areas
- over 400 educational programs from SberUniversity for professional and career development
- adaptation program and supervisor support at the start
- extended voluntary medical insurance, preferential insurance for family, and corporate pension program
- flexible mortgage discount equal to 1/3 of the Central Bank's key rate
- free Prime subscription with the option to share with three close contacts
- reward for recommending friends to the Sber team.
