Looking for a job

Anton Koptev

@antonkoptev

Senior AI Engineer at Wildberries Tech building a multi-agent assistant and owning Smart Search, a production RAG system for sellers.

Georgia

Message

At Wildberries Tech, I’m building an internal multi-agent assistant and own Smart Search, a RAG system for sellers. The assistant supports SQL reports, text-to-SQL, help-center search, and PDFs; I designed its stateful orchestration and wrote its router, agents, and tools.

I built the evaluation pipeline for the agents, checking routing and tool calls against a golden set and call contracts. A Gemma judge grades Qwen agents’ trajectories and answers, while production traces and user feedback feed regression cases and future agentic RL.

Smart Search covers about 8,000 help articles. It combines BGE-M3 and BM25 retrieval with cross-encoder reranking, cites source sections, and returns a full answer in about three seconds at p95; I also added blue-green index swaps in Qdrant and search observability.

At CDEK Digital, I built an LLM support assistant that cut cost per request by 27% and operator workload by 35%, and led three DS/ML engineers while continuing to code. Earlier at POWERTEC, I built contextual-bandit auction optimization that increased revenue per thousand impressions by 4–9%.

Experience

Work history, roles, and key accomplishments

Current

Senior AI / ML Engineer

Current

Wildberries Tech

Sep 2025 - Present (1 year 1 month)

Internal multi-agent assistant for employees and Smart Search, a production RAG system for sellers that I own end-to-end.
• I architected the assistant's stateful orchestration and wrote the router, the agents and their tools. The router is LLM-based; tool chains are deterministic and each step is traced. In the decision layer, TypeSafe Jev scores the allowed next actions as probabilities, and the

Senior AI Engineer / Data Scientist

CDEK Digital

Sep 2023 - Aug 2025 (1 year 11 months)

LLM assistant and ML systems for customer support and customs operations at CDEK, a large logistics company.
• Built an LLM assistant for first-line support: hybrid BM25 and dense retrieval with reranking, orchestrated in LangGraph, served with vLLM on KServe. It raised Precision@5 by 19 pp and cut cost per request by 27% and operator workload by 35%.
• For customs, built HS (tariff) code classifi

Machine Learning Engineer / Data Scientist

POWERTEC

Jun 2021 - Sep 2023 (2 years 3 months)

Applied ML for real-time advertising: auction optimization and brand safety.
• Built RTB auction optimization with contextual bandits, dynamic floor pricing via Thompson Sampling and LinUCB: revenue per thousand impressions rose 4–9% and auction timeouts fell 12–18%.
• Built brand-safety semantic matching on E5/BGE embeddings, with ANN retrieval and reranking. Unsafe impressions fell by 20–28%, ma