Kinetic Sunyaev-Zeldovich (kSZ) effect is a secondary CMB anisotropy and an interesting cosmological probe that is going to be well measured soon. We focus on the application of kSZ to the reconstruction of cosmic radial velocity. We develop an optimization based likelihood approach for the kSZ-induced cross-correlation of CMB with the large scale structure via kSZ effect. We show that while existing quadratic estimator technique is enough for near-term experiments, the proposed likelihood approach can give higher signal-to-noise in the future. The autodifferentiable likelihood is implemented and optmized in JAX and evaluated on Agora simulations. We further discuss possible extensions and improvements, such as machine-learning based estimator of electron density given an observed galaxy distribution.