Correct, the Builder tools are for crystalline, well-behaved cases where the coincident lattice model makes sense.
There is not really a simple algorithm to create an interface between two random materials. At heart it's a global optimization problem, hard in itself, and any starting point (the configurations you happen to have of the left and right material, respectively) it's just a representative case, and can never be said to be universal, since each time you generate the amorphous sample you will get a new random configuration, and how you choose the surface termination is equally random.
Now, that may have been obvious to you already, but I just wanted to ensure we share the same picture.
As for the approaches you mention, I would definitely try the universal NN-based potentials for this case. They may not be perfect, and often not accurate enough for a reaction barrier or other very detailed calculations, but the time you save from not having to create a new MTP from scratch or even a fine-tuned MACE demo model from compensates for this. And, you can (and probably should) always do a final DFT relaxation - that is, the MLP-generated structure should just be seen as a starting guess for the interface configuration. And, thanks to their speed and readiness, you can try generate dozens or hundred of candidates in very short time.
