Vlad Tsybin

I'm an AI/ML Engineer with 4+ years of experience, focused on production LLM applications: AI agents, RAG and LLM evaluation. I care about measurable quality and debugging from traces rather than guessing.

In my current role I designed and built a production AI agent framework from scratch (FastAPI): custom tool selection, Model Context Protocol (MCP) as both client and server, and an adaptive RAG pipeline where an LLM query planner routes between lightweight and deep retrieval. I built the retrieval pipeline (document parsing with Docling, chunking, BGE-M3 embeddings, vector search, reranking) and raised relevant-chunk recall from 70% to 95%.

I build evaluation and observability around LLM systems. Langfuse tracing plus custom tool-use metrics cut issue detection and debugging time by 50%. Frozen evaluation sets let me compare prompt variants for multi-area RAG before it went to production. Prompt engineering, model selection and fallback logic raised correct tool usage by 4 percentage points. I also built a PII masking pipeline (regex + NER + LLM) that raised masking F1 from 0.66 to 0.92.

Before that, I built RAG assistants with tool calling over an OLAP system, compared open-source LLMs (Qwen, DeepSeek), and shipped demand forecasting models (CatBoost, Prophet, SARIMA) into production for a major quick-service restaurant chain.

Completing an MSc in Artificial Intelligence at ITMO University (2027).

Stack: Python, LangGraph, LangChain, OpenAI Agents SDK, FastAPI, asyncio, vLLM, PyTorch, FAISS, Langfuse, Docker, SQL.