Errors in in situ stresses narrow the mud-weight window and compromise wellbore stability and hydraulic fracture containment, yet remains the most poorly constrained stress component. Wellbore deformation relative to a known reference configuration can constrain the stresses, but conventional caliper and image logs measure borehole shape rather than the material displacement field required by the inversion; analytical inversion based on the Kirsch solution assumes a circular borehole in a homogeneous elastic medium. Finite element modeling offers a route beyond this restriction, but gradient-free optimizers require thousands of forward solves. We present INSIDE (IN Situ stress Inversion from wellbore DEformation), which couples a GPU-accelerated finite element forward model with L-BFGS-B optimization to recover the far-field stresses and from wellbore wall deformation. While the problem has only two unknowns and each finite-difference gradient requires only three forward solves, the iterative workflow is deliberately retained as a general nonlinear inversion path for stress-dependent constitutive models. The forward model reproduces the Kirsch solution with wall-displacement errors below 0.24%, while noise-free twin experiments recover both stresses with relative errors below 0.01%. Monte Carlo ensembles of 100 seeds per noise level give mean stress errors of 1.60% () and 2.18% () at 20% noise, converging within 4–6 iterations regardless of initial guess. Young’s modulus is the dominant error channel, with stress errors scaling nearly one-to-one with modulus uncertainty. This study deliberately focuses on a pure-elastic baseline in a circular, homogeneous, and isotropic medium. Extending this framework to non-circular geometries, heterogeneous media, and elastoplastic behavior represents the primary direction for future work.