Customized vision sensors for robotic welding demand rapid re-targeting of optical, mechanical and electronic parameters, yet conventional CAD/CAE loops still require weeks of expert effort. We present an end-to-end, data-driven framework, that couples high-fidelity ray-tracing and thermo-mechanical simulation with surrogate neural models and multi-objective Bayesian optimization to automatically design and fabricate SkyBlue cameras tailored to any welding task. The pipeline explores a 25-dimensional design space (field-of-view, working distance, housing stiffness, sensor choice, etc.), converging on manufacturable solutions that jointly maximize imaging precision and minimize cost, weight and lead-time. For two contrasting variants—a high-precision 30 mm model and a wide-view 120 mm model—the approach cut design iterations from 12 to 4, reduced prototype cycle time from six weeks to two, and lowered bill-of-materials cost by 29%, while predicting modulation-transfer-function within 3% of laboratory measurements. These results offer a transferable blue-print for first-time-right production of bespoke optical devices.