Skip to content

Александр Бояхчян

@alexbronte

AI / LLM ENGINEER | PRODUCTION LLM SYSTEMS, AGENTS & INFERENCE

I keep finding my way back to the practical side of language models, from improving search over thousands of documents to getting an agent to check subscriptions. Lately, I have been running models on a two-GPU server.

Career

Experience: 4 years

Experience

AI / LLM Engineer

League of Digital Economy

Russian IT company that develops and implements digital solutions for large businesses, including software and technology platforms.

Aug 2024 — Present

Built and deployed a tool-using support agent, resolving about 40% of routine requests without an operator and reducing first-line workload by 35-40%.

• Built a stateful LangGraph support agent that checks payment and subscription information through internal APIs, handles routine requests, and hands unresolved cases to an operator. • Connected internal tools through MCP. Added Pydantic validation, access checks, duplicate-call protection, and result verification for actions that change data. • Evaluated 650 scenarios, reaching 91% tool-selection accuracy and 87% end-to-end task completion. Around 40% of routine requests were resolved without an operator, reducing first-line workload by 35-40%.

AI / LLM Engineer

Aspirity

Software development and IT services company that builds custom digital products and technology solutions for business clients.

Jan 2022 — Jul 2024

Built a RAG assistant and support-message classifier, improving retrieval metrics, cutting search time, and reducing routing errors to 4-5%.

• Built ingestion and index-update workflows for Confluence and PDF sources, including cleaning, deduplication, table handling, and OCR. The assistant served about 150 active users across a corpus of approximately 9,000 documents. • Combined BM25 and Qdrant dense retrieval with Reciprocal Rank Fusion (RRF) and cross-encoder reranking. Improved Recall@5 from 0.83 to 0.91 and MRR@10 from 0.62 to 0.74. Search time fell from 3-5 minutes to 40-60 seconds, with 18% fewer second-line support requests.

Education

Bachelor's degree, Applied Mathematics

Novosibirsk State University

2016 — 2020

Skills

vLLMGPU inferenceQwenTransformersLoRAtool callingstructured outputhuman handoffLangGraphMCPQdrantBM25dense retrievalRRFrerankingembeddingsRAGASPythonPyTorchFastAPI