Developed an end-to-end evaluation infrastructure for case-level faithfulness within the NVIDIA-NeMo/Gym repository, focusing on automated assessment of model outputs across QA, summarization, and data-to-text tasks. Built the RAGTruth Resources Server using Python, FastAPI, and Pydantic, enabling dataset preparation, verification scoring, and comprehensive unit testing. Implemented robust parsing logic with regex to handle noisy model outputs, including the removal of extraneous tags and JSON fences. The system reports accuracy and F1 metrics and can operate as a standalone scorer without requiring a full model server, supporting flexible deployment in air-gapped environments and facilitating reproducible, automated evaluation workflows.