We propose a novel method for automated short answer grading (ASAG) designed for practical use in real-world settings. The method combines LLM embedding similarity with a nonlinear regression function, enabling accurate prediction from a small number of expert-graded responses. In this use case, a grader manually assesses a few responses, while the remainder are scored automatically--a common scenario when graders need to review some responses to feel confident assigning final grades. The proposed method achieves an RMSE of 0.717 outperforming the fine-tuned state-of-the-art transformer models in grading accuracy, which are more labor-intensive and computationally demanding, limiting their practicality for many applications. This method stands out for its ease of implementation and effectiveness, offering reliable accuracy with minimal effort. The code is made public. [For the complete proceedings, see ED675583.]