In December 2025, Max Tennenhaus enhanced compute time estimation for batch processing in the ai-dynamo/nixl repository by integrating tensor parallelism scaling and detailed MLP FLOPs calculations. Using Python and leveraging machine learning expertise, Max improved the accuracy of model inference benchmarks, which supports more reliable capacity planning for multi-GPU deployments. The technical approach involved refining performance metrics by modeling both attention and MLP layer FLOPs, and collaborating with NVIDIA engineers to align KVBench with these new methodologies. This work demonstrated depth in performance optimization, addressing ambiguity in existing formulas and enabling more precise benchmarking for large-scale inference workloads.