This article introduces a novel in-processing method for integrating actuarial and equity fairness into neural networks used for actuarial valuation. We consider one primary network penalized during training to ensure balanced predictions (actuarial fairness) and independence from sensitive features (equity fairness). Global and local actuarial equilibrium is obtained by aligning the inter-quantile averages of predicted and observed responses. Meanwhile, a second auxiliary network penalizes the primary network for discriminatory predictions. The combined training algorithm effectively preserves predictive accuracy while mitigating discrimination. Numerical illustrations on real-world datasets demonstrate the method’s efficacy in achieving fair and reliable insurance pricing models.