Nikita Marulin

@nikitapython

Senior AI/ML Engineer

NLP Engineer / ML Engineer with 6+ years of commercial development experience.

About

I build and ship end-to-end NLP & LLM solutions for production - from agentic AI assistants that take real actions across enterprise systems to multi-tenant RAG pipelines and high-scale NLP systems for fraud detection, search and content moderation. I know how to turn business goals into ML tasks, collect and label data, train and fine-tune models, and keep the whole pipeline healthy and measurable in production. I'm genuinely passionate about applied AI and how agentic systems are reshaping how businesses operate.

Career

Experience: 6 years

Experience

Senior NLP Engineer

Intercom

Customer service software company offering messaging and support products.

Jun 2023 — Present

Built agentic actions, knowledge-base auditing, and RAG retrieval for Intercom's AI support agent; improved resolution from 31% to 58%, with zero irreversible action errors.

• Built an agentic layer that shifted the AI support assistant from text-only replies to executing actions across tenant systems, using a ReAct loop on LangGraph over a tool registry with argument contracts and GPT-4o function calling. Added a deterministic guardrail for operation reversibility and confirmations, reaching tool selection accuracy of 0.91, argument accuracy of 0.87, and stopping accuracy of 0.94. Resolution rate rose from 49% to 58% with zero irreversible errors.

ML Engineer

Vinted

Online marketplace focused on secondhand fashion and other pre-owned goods.

Apr 2020 — May 2023

Developed ML systems for marketplace fraud detection, search, moderation, and attribute extraction; reduced fraud tickets by 38%, accelerated takedowns, and improved search and listing conversion.

• Built a scam-detection classifier for marketplace chat that replaced 340 hand-written rules (precision 0.44), combining an XLM-RoBERTa branch with a deobfuscation layer and a character-level CNN on raw text. Obfuscation-based augmentation added 4.7 points of recall, reaching final precision of 0.87 and recall of 0.71 while cutting false positives from 8% to 1.2%. Fraud-related support tickets dropped 38%, and the model is now used by the Trust & Safety team.

Education

Master’s Degree, Computer Science and Computer Engineering

Moscow Institute of Electronic Technology (MIET)

2023

Skills

PythonSQLBashLLMRAGLangChainLangGraphTransformersFine-tuning (LoRA / PEFT)AI AgentsPrompt EngineeringLLamaQwenMistralOpenaiXGBoostClusteringUpliftNLPCatBoost