Looking for a job

Никита Марулин

@0007068

Message

Senior NLP Engineer at Intercom who raised AI support resolution to 58% with agentic actions and production RAG.

Brazil

Message

At Intercom, I built an agentic layer that lets the AI support assistant take actions across tenant systems, raising resolution from 49% to 58% with zero irreversible errors. I also designed retrieval for a multi-tenant RAG system, improving resolution from 31% to 49%.

At Vinted, I built NLP and machine learning systems for scam detection, search, listing moderation, and attribute extraction. My work helped reduce fraud-related support tickets by 38% and cut counterfeit takedown time from 4.1 days to 3.5 hours.

Experience

Work history, roles, and key accomplishments

Current

Senior NLP Engineer

Current

Intercom

Jun 2023 - Present (3 years 4 months)

Built an agentic layer for the AI support assistant using ReAct on LangGraph with GPT-4o function calling, improving tool selection accuracy to 0.91 and resolution rate from 49% to 58%. Developed a knowledge base audit pipeline and retrieval core for a multi-tenant RAG system, raising faithfulness to 0.91 and Recall@10 from 0.68 to 0.90.

LangGraph GPT 4o React RAG BM25 Elasticsearch BGE M3 Qdrant RRF DeBERTa V3 NLI UMAP HDBSCAN Fine Tuning Python

ML Engineer

Vinted

Apr 2020 - May 2023 (3 years 1 month)

Built a scam-detection classifier for marketplace chat, replacing 340 hand-written rules and achieving precision of 0.87 and recall of 0.71. Rebuilt marketplace search retrieval with BM25 and FAISS via RRF fusion, and automated listing moderation across 23 categories, reducing takedown time from 4.1 days to 3.5 hours.