Description
Sberbank Online (SBOL) — applications with 80+ million clients. In the application, in addition to the familiar and classic functionality, we are developing AI services, the main one of which is the AI Assistant.
The main challenge for us is to use the familiar client search in SBOL, supplement it with the functionality of the AI Assistant so that, without losing the quality of basic functions, we can learn to solve more complex client tasks.
With the participation of the Lead Data Scientist, we want:
- develop the DS competence at the unit level. This includes working with those teams that are engaged in or are preparing to engage with agents or functions for the AI Assistant.
- define the development vector for Search with AI assistant in SBOL, jointly with other teams. This direction is largely related to adapting LLMs to the specifics of the product and to assessing the quality of Search and the AI Assistant — and then to the changes that will need to be made based on the assessment results. On the one hand, this is data research and hypothesis testing, and on the other — building a common and transparent process.
Responsibilities
LLM-oriented solutions:
- optimizing LLM performance in production (speed, cost, accuracy)
- designing and developing pipelines for data processing (RAG, agent systems, semantic search).
Working with data:
- data processing pipelines
- setting requirements for labeling
- dataset quality control
- evaluating ROI of ML features through metrics / experiments.
Leadership and expertise:
- ability to translate business goals into ML tasks, formulate requirements for DS teams, participate in product and team development strategy
- participation in setting technical requirements and interacting with business customers
- working out requirements and solution options with product managers, system analysts, and related teams
- working out test and training labeling for teaching legal skills to GigaChat and other LLMs with the training department
- selecting and developing people, mentoring junior colleagues, developing best practices for the team
- analyzing risks and finding compromises between model quality, speed, and cost.
Production engineering:
- implementing DS-models in production using MLOps practices. (CI/CD monitoring, A/B tests).
Requirements
- higher education, work experience from 5 years in DS/NLP, including at least 1 year of work with LLMs, experience with production
- willingness to both write code, pipelines, train models, and write documentation, design systems, and prepare specifications for models, data, pipelines
- deep expertise in adapting LLMs: SFT, RLHF, LoRA, prompt engineering
- experience in building RAG systems, agent pipelines and services based on LLMs, knowledge of modern frameworks (PyTorch, Hugging Face, LangChain, LlamaIndex)
- understanding of infrastructure components: Docker, Kubernetes, cloud platforms, MLOps: CI/CD, data drift monitoring, logging
- experience in transforming business tasks into technical requirements
- ability to evaluate ROI of DS solutions and balance between innovation and practicality, ability to quickly prototype solutions and find a balance between speed/quality/performance.
Conditions
- work format: office in Moscow, hybrid - by agreement, Business Center, Kutuzovsky Prospekt, 32, building 1
- annual salary review and annual bonus
- more than 400 educational programs from SberUniversity for professional and career development
- extended VHI (voluntary health insurance), preferential insurance for family
- free SberPrime+ subscription, discounts on products from partner companies
- referral bonus for recommending friends to the Sber team.
