Towards deployment-centric multimodal AI beyond vision and language