Design a modular pipeline for a perception system at ABB. How do you ensure the pre-processing step doesn't bottleneck model inference?
