September 28, 2026 • Moscow • PJSC Sberbank • Information Technology: Data Scientist
No experience required
The team is developing NLP models and LLM-based solutions for operational risks. We are solving problems related to building AI agents and multi-agent systems.
Examples:
- Gardrails system (including on Jev-like models) to protect Sber's LLMs and AI agents from prompt attacks and jailbreaks
- Bank's agent memory based on a temporal knowledge graph that considers versions and changes in internal documentation over time
- Harness that autonomously transforms DS code into a production service — from project assembly according to standards and integrations to deployment on test stands and testing
- AI agents for assessing the risks of Sber's ecosystem companies based on internal documents, news, reports, and business process descriptions
- Multi-agent system for analyzing code repositories for business logic vulnerabilities and checking against requirements
- AI Data Engineer agent for autonomous data analysis, error and anomaly detection, creation of data quality checks, and creation of extracts and data marts.
Responsibilities
- Development of PoC and MVP AI agents using frameworks for working with LLMs
- Adaptation and fine-tuning of LLMs on internal and external data (SFT, LoRA, DPO, GRPO)
- Participation in deploying AI agents to production, their integration, and quality monitoring
- Writing high-quality code and discussing technical solutions with the team.
Requirements
- Incomplete higher technical education with a strong background in mathematics and algorithms
- Knowledge of Python, statistics, and the ML/DL stack
- Experience in designing and training neural networks: from CNN/RNN to Transformers
- Experience building RAG systems (classic RAG, GraphRAG, Agentic RAG)
- Experience developing autonomous AI agents (LangChain, LangGraph, Deep Agents)
- Knowledge of classic algorithms and data structures, OOP principles, and design patterns
- Ability to write clean and maintainable code (Clean Architecture, SOLID)
- Experience with Git.
Will be a plus
- Experience with FastAPI, Docker, vLLM, Unsloth, Langfuse
- Success in competitions (e.g., Kaggle Expert+ or prize-winning places at hackathons)
- Contributions to open-source projects, publication of models and datasets on Hugging Face
- Deep understanding of industry trends, interest in recent articles and new tools
- Knowledge of other programming languages (e.g., C++, Java).
Conditions
- Paid internship
- Option to choose a convenient schedule – office or hybrid (location: Kutuzovskaya metro station)
- Corporate gym and recreation areas.
