Roman Shaposhnikov

I designed Uplift's WhatsApp AI assistant architecture for exhibitor, schedule, and sponsor questions, combining hybrid retrieval, confidence gating, cost controls, and two-layer hallucination checks.

At Yesbat, I led modeling for a battery state-of-health predictor across 793 cells and nine public dataset families, achieving 0.025 average SOH error on the certified cohort. I built reliability gates to withhold unsupported predictions and corrected misleading validation reporting through a metric-honesty audit.

My PhD and postdoctoral work at the University of Pavia and University of Milano-Bicocca built reproducible Python, R, and Bash pipelines for genome-scale sequencing data on Linux/Slurm HPC.