Data Scientist (NLP|LLM)

Moscow Company: SberZdorovye ☑️What will you be doing? -Designing and supporting the full cycle of medical LLM improvement: data collection, cleaning, versioning, training, and fine-tuning (SFT, preference-tuning, DPO/ORPO, instruction tuning). -Building datasets and annotation pipelines: schemes and guidelines, consistency control, synthetic data generation, self-training, error and bias analysis. -Developing LLM-based pipelines and agents for medical tasks: RAG on clinical guidelines and reference books, tool-calling, routing, multi-step workflows, orchestration (LangGraph and multi-agent frameworks), guardrails. -Creating and developing an evaluation system: test sets and benchmarks, automatic metrics and LLM-as-a-judge where appropriate, expert validation with doctors, red-teaming, regression runs, A/B testing in production. -Conducting research iterations: formulating hypotheses, setting up experiments, performing ablation studies, documenting results, preparing scientific articles, and bringing materials to publication.

☑️Required Technical Skills: -3+ years in NLP/ML, proficient Python: typing, testing, profiling, clean production code. -Practical experience in training and fine-tuning transformers: PyTorch + HuggingFace, understanding of Accelerate, DeepSpeed, or similar. -Experience in building data pipelines and reproducible experiments: datasets, versions, configs, tracking (MLflow or ClearML), ability to make correct comparisons. -Understanding of LLM systems: retrieval, tool-calling, agents, quality degradation, hallucinations, production limitations. -Quality evaluation skills: metrics, benchmarks, error analysis, ablations, working with annotation and expert validation.

☑️Will be a plus -Experience in medicine or biomedicine: clinical texts, ICD-10, clinical guidelines, understanding of domain risks. -Experience in alignment: preference data, RLHF, DPO, safety eval, hallucination mitigation. -Practice in production inference: vLLM, cost and latency optimization, Docker, K8s, monitoring, tracing. -Experience with retrieval stack: hybrid search, rerankers, FAISS, Elastic, pgvector, thoughtful chunking, grounding.

☑️What we can offer: -A strong team of professionals passionate about their work; -Opportunities for development within a leading MedTech company in Russia; -A cozy office in the City with panoramic city views, hybrid work format; -Company equipment; -Medical program including telemedicine consultations, in-person appointments at clinics, psychologists, dentistry, laboratory and instrumental diagnostics; -Paid English language courses; -We support an active lifestyle — choose sports you like (corporate squash, running, football in Moscow and compensation for your sports membership); -SberUniversity and payment for specialized training and courses. Contacts: recruiter contact