We trained an MLP that predicts 58 body shape parameters from 8 questionnaire inputs. The trick: a differentiable physics loss. Mass accuracy: 0.3 kg mean vs 3.9 kg for Ridge. No GPU at inference.