Mosoblgaz
AI Engineer / LLM Engineer
from 235,000 ₽ per month, net
Experience: 3–6 years
Full-time
Schedule: 5/2
Work format: hybrid or office (Moscow)
☑️ Responsibilities
- Develop agents on LangGraph: tool calling, multi-step scenarios, human-in-the-loop, error handling
- Build RAG on internal documents: parsing PDF/DOCX, scans and tables, indexing, hybrid search (BM25 + dense), reranking
- Create and develop agents for corporate systems (Bitrix, Jira, Alfresco, Exchange, 1C): text-to-SQL for MS SQL / PostgreSQL, integrations via REST API
- Analyze agent errors by traces, maintain regression sets and quality metrics
- Work with customers to take solutions from prototype to production and document them
☑️ Our requirements for candidates
- From 2 years of Python development, including at least 1 year on LLM projects (agents, assistants, RAG). asyncio, FastAPI, pydantic, tests
- LangGraph, LangChain or similar; function calling, structured output, agent patterns (ReAct, plan-and-execute, routing)
- Practical RAG experience: chunking, embeddings, vector and hybrid search (Qdrant, pgvector or similar)
- Understanding of LLM limitations and how to overcome them: context window, hallucinations, non-determinism. Ability to determine when an LLM is needed, when a deterministic solution is sufficient, and when classic ML is appropriate, and to justify the choice based on quality, cost, and reliability
- Quality assessment: assemble a test set, choose metrics (recall@k, faithfulness, RAGAS or similar), conduct offline evaluation and side-by-side comparison of prompt or pipeline versions
- Confident SQL; Docker / Docker Swarm, GitLab CI
- Ability to communicate with business customers, translate tasks into technical specifications, and write clear documentation
- Bonus points:
- Work with open-weight models (Qwen, DeepSeek) and their deployment on vLLM: quantization (FP8/AWQ), guided decoding
- Experience with Russian LLMs (GigaChat, YandexGPT) via API
- LiteLLM as a gateway to models; OpenWebUI (Functions, Tools, Pipelines)
- Langfuse or LangSmith, LLM-as-a-judge
- MCP: creation and connection of servers
- LLM system security: prompt injection, masking of confidential data
- Classic ML for text: classification, clustering, boosting, sklearn (e.g., for request routing)
Contacts: recruiter contact
