The l1-penalizing method introduced in "Learning Multiple Quantiles With Neural Networks" provides non-crossing quantiles estimates of response variable for given explanatory data.
@article{moon2021learning,
title={Learning multiple quantiles with neural networks},
author={Moon, Sang Jun and Jeon, Jong-June and Lee, Jason Sang Hun and Kim, Yongdai},
journal={Journal of Computational and Graphical Statistics},
volume={30},
number={4},
pages={1238--1248},
year={2021},
publisher={Taylor \& Francis}
}