Developed multi-metric support for the RecMetric class in the pytorch/torchrec repository, focusing on enhancing evaluation capabilities for recommendation systems. The work involved refactoring the handling of predictions, labels, and weights to enable robust computation across metrics such as CTR and NE. By streamlining the metric computation pipeline, the changes allowed for per-metric aggregation and improved reusability of RecMetric, supporting faster experimentation with evaluation strategies. Implemented primarily in Python with a focus on data analysis and machine learning, this feature improved evaluation fidelity and decision-making for CTR optimization, while maintaining tight integration with the existing TorchRec codebase.