Worked on enhancing batched inference robustness in the EvolvingLMMs-Lab/lmms-eval repository by implementing per-request sampling parameter isolation for VLLM batch processing. Addressed the challenge of preserving unique sampling parameters for each request within a batch, ensuring that parameter leakage was prevented and output consistency was maintained. Utilized backend development and concurrent programming skills in Python to deliver a feature that supports more personalized and reliable outputs during batched deployments. Focused on unit testing to validate the preservation of sampling parameters, aligning the solution with product goals for improved governance and tailored inference results in large-scale machine learning experiments.