Y'know, I haven't personally tried it, but there's a pretty good chance that ending most closed hardware and software is just as far away as training transformer models to automatically reverse them.  The biggest impetuses to that are collecting example data and setting up the training environment.  The best example data might be from chip design simulators if such exist, as then you could autogenerate variations and also automatically change points at which voltages are probed.