I think @SEJPM's suggestion is a solid way to accomplish the task if you wanted a custom feistel network that probably also works :P
If you're just having fun, or if you feel like working very hard and don't have any heavy risk involved, then you might be able to apply the leftover hash lemma to something like a subset-sum instance
I think subset-sum is even NP-complete, but no surprise here: it's riddled with cases that admit easier solutions than brute force
For example I believe anything with "density" away from $1$ on either side is trivially solved by lattice reduction, but still it would be a shock if a poly time algorithm covered every case
Lets see, I think the syndrome of a random parity check matrix is considered a "good" hash function in some respect