I'd go for some custom pixel shader with depth testing...

Maybe there's some other, clever way using ZFunc and rendering the model several times.

EDIT: yeah, writing the model with a first pass (blue-see-through-wall-colored, no depthtesting, no alphawrite) and then rendering the model in a second pass as usual (with depthtesting and alphawrite)

maybe it's even easier, I don't know what kind of stuff zFunc supports since I never use it...