We investigate buffer-aware demand-routing of key-consumption demands in QKD networks and implement a measurement-based framework for path selection. The proposed ML-based algorithm outperforms state-of-the-art heuristics by 22 % and 92 % for networks under test.