At Analytical Programming Solutions, I contributed to a production system matching customs declarations to 2,600+ regulatory documents. It used BERT classifiers and multi-stage pre-filtering, while my R&D included synthetic training data and fine-tuned LLM2Vec retrievers.
For Gazprom, I co-developed an on-premise pipeline to flag contradictions between project documentation and regulations. I also built RAG assistants over document sets and designed an extraction prompt for construction scheduling.
On an independent project, I built an LLM support agent for an ad-tech platform, combining hybrid retrieval with tools for SQL queries and log search. I added safeguards including SQL validation and prompt-injection filters.
My first-author paper on retrieval-augmented terminology translation for English–Russian was accepted to the WMT 2026 Research Track. The work explores a multi-domain RAG pipeline and fine-tuned translation models.
