Our team works on machine learning for Yandex Maps. We help users find the places they want to get to and decide where to spend their time. We also work with user-generated content, such as reviews and photos, to highlight the key features of businesses. Furthermore, we digitize the real world: we automatically build the map from satellite images, panoramas, and photos.

Using ML, we strive to make our approaches scalable and solve tasks that previously only humans could handle. For example, recommending good places for very complex queries or automatically describing a place a user is interested in.

In our team, we conduct full-stack development—from the idea for solving a task to its deployment in production. We analyze issues in model responses, verify the performance of methods by studying offline and online quality metrics, and implement a trained model to work in runtime or in a regular pipeline.

Our tech stack: Python, C++, CatBoost, BERT, YaLM (Zeliboba), YandexGPT, YTsaurus, segmentation, detection, vectorization models, and NeRF for CV tasks.

About the team

https://yandex.ru/project/vacancies/ml_analytics_maps

What tasks await you

Search quality How do you find stores where you can buy a rare car part? Which beauty salon dyes hair in a special way? Answering such questions is not easy, and for an ML developer on our team, it is a separate challenge. We improve the quality of search results for complex queries, relying on all the information we know about a business, including reviews, photos, websites, and information about menus and services.

Place selection and personalization We launched a new "Ideas" mode in Maps. In it, we suggest to users where to go and what to do, and tell them about these places by summarizing vast amounts of information about each place using YandexGPT and computer vision models. Our team works on improving the quality of recommendations so that the selected businesses match the user's wishes as closely as possible.

Large language models for training other neural networks We use a lot of assessor labeling to train models. One way to increase assessor efficiency and the usefulness of training data is to assign only complex cases to humans for labeling, and apply heavy LLMs for simple decisions. Our team increases the efficiency of label collection and the overall quality of training data by using YandexGPT, ultimately aiming to improve the user experience by offering more accurate search ranking, more suitable suggestions, and place descriptions.

We expect that you

• Have a solid understanding of classical ML, RecSys, NLP, or CV principles
• Know C++, Python, or another programming language
• Have knowledge of fundamental algorithms and data structures