Rigers Aliaj

At the University of Hamburg and DESY, I develop graph-neural-network models for scientific computing problems where computational cost matters. Using more than 20 million scientific graph samples, I translated an open-ended research problem into an ML filtering system for an expensive downstream sparse linear-system calculation.

I've benchmarked classical baselines, GIN, GAT and Graph Transformers in PyTorch and PyTorch Geometric, and built reproducible GPU/CPU workflows with Slurm and Weights & Biases. At the selected operating threshold, my work removed approximately 64% of candidate graphs and reduced the downstream system by approximately 30%.

I also independently built a Python and pandas payment-matching workflow for Bestherm, turning days of manual reconciliation into minutes for a firm serving approximately 200 customers.

My mathematical physics background helps me identify structure, test assumptions and balance model sophistication against computational cost. I enjoy moving from unclear quantitative questions to documented, measurable implementations across physics and machine-learning contexts.