RocketHuntSmart job search
Reach out directly about this role
By job title
3-5 years
Experience
Full-time
Employment
Hybrid, Onsite
Work Format
Data Science & ML
Specialization
IT & Tech
Industry
Corporation
Company Type
Yandex Maps GeoSearch helps millions of people around the world find the places they need every day. We answer a wide variety of queries: from the simple "тамақтану" in Kazakh to the complex "Boğaz manzaralı hoş bir restoran" in Turkish.
Our main task is to understand the user's intent in fractions of a second, regardless of the language they write their query in or where in the world they are, and to show on the map exactly what they are looking for. We transform a text query into a geographical answer, taking into account a huge number of factors: from traffic and opening hours to cultural specifics and implicit preferences.
Maps are a core technology for many of Yandex's urban services, such as Yandex Taxi, Yandex Delivery, and Yandex Food.
Why this is an interesting and challenging task * Working in international geo-search is a unique challenge at the intersection of technology, language, and culture. Unlike classic web search, we deal with the physical world, and that changes everything. * The language barrier. A user may search in Russian in Turkey, in English in Armenia, or use transliteration. We need to understand queries in different languages, mixed queries (where one query contains several languages), and handle the variety of names for the same place (Eiffel Tower, Эйфелева башня, tour Eiffel). * Cultural and local specifics. The concept of "city center," "popular place," or "best coffee" in Istanbul, Yerevan, or Dubai are three different things. Our models must adapt to the local context and understand what is important for the user in a specific country. * The multifaceted nature of user needs. The same query "Apple" can mean an electronics store, a company office, or a fruit market. Our algorithms must determine the correct context in real time, based on geolocation, time, and user history. * Scale and speed. All this complex understanding must work with minimal latency for millions of users simultaneously. This requires not only smart but also incredibly efficient models.
You will be working on a product where the result of your code instantly impacts the experience of millions of people, helping them navigate unfamiliar cities and discover the world.
Development of multilingual NLP models You will teach our search to understand queries in different languages, extract entities from them (what is being searched for, where, with what attributes), and handle typos and informal formulations.
Improving search ranking for different countries You will create and train classical and neural network models that will rank top organizations taking into account local specifics: popularity, reviews in the local language, transport accessibility, and dozens of other factors.
A/B testing and model deployment You will need to develop pipelines for quickly testing hypotheses and rolling into production only those solutions that have proven their effectiveness on real users.
Building vector representations We use state-of-the-art architectures to create a unified semantic space where the query "a cozy place for dinner" and the Turkish name of a restaurant with corresponding reviews will be close together. This allows us to find relevant places even if their description lacks direct keywords.
Generative approaches We are exploring how generative models can help in understanding complex, detailed queries ("looking for an inexpensive cafe where I can bring my dog and which has power outlets for a laptop") and even in forming personalized recommendation responses for the user.
More about ML at Yandex — in the channel Yandex for ML