An observer is an estimator of the state of a dynamical system from noisy sensor measurements. The need for observers is ubiquitous, with applications in fields ranging from engineering to biology to economics. The most widely used observer is the Kalman filter, which is known to be the optimal estimator of the state when the noise is additive and Gaussian. Because its performance is limited by the sensors to which it is paired, it is natural to seek an optimal sensor for the Kalman filter. The problem is however not convex and, as a consequence, many ad hoc methods have been used over the years to design sensors. We show in this paper how to characterize and obtain the optimal sensor for the Kalman filter. Precisely, we exhibit a positive definite operator which optimal sensors have to commute with. We furthermore provide a gradient flow to find optimal sensors, and prove the convergence of this gradient flow to the unique minimum in a broad range of applications. This optimal sensor yields the lowest possible estimation error for measurements with a fixed signal to noise ratio. The results presented here also apply to the dual problem of optimal actuator design.