Daniel Dragonevskiy

My research sits at the intersection of ML, LLMs and computer vision, with a focus on inference optimization (KV-cache compression, quantization) and ensemble methods for large-scale forecasting. In my recent paper (IJCISIM 2026, DOI: 10.70917/ijcisim-2026-4333) I showed analytically and empirically that geometric log-space ensembling is the MAPE-optimal aggregator for log-normal targets, reaching MAPE ≈ 4.77% on an 18,000-store retail forecasting challenge. Open directions: efficient LLM inference, vision-language models, and transferring multiplicative-error ensembling to new domains