Papers would be good next step. The concept and the algorithms do not seem very complex. The training and fine tuning is what needs effort and infrastructure.