This study aims to evaluate the effectiveness of the grid-search method in hyperparameter optimization on Teachable Machine (TM) using a varying number of image samples. The hyperparameters studied include epoch (e), batch size (b), and learning rate (l). A structured grid-search method approach will be applied to test 216 hyperparameter combinations across 6 categories of sample size per class, namely 10, 25, 50, 100, 250, and 500. The results showed that the optimal combination findings were obtained based on variations in the number of samples as follows: 10 samples using e:100, b:256, l:0.001 get an accuracy range of ≥ 90%; for 25 samples using e:500, b:16, l:0.001 get an accuracy range ≥ 97%; for 50 samples using e:100, b:512, l:0.001 get an accuracy range ≥ 88%; for 100 samples using e:500, b:32, l:0.001 get an accuracy range ≥ 88%; for 250 samples using e:50, b:16, l:0.001 get an accuracy range ≥ 92%, and finally 500 samples using e:500, b:256, l:0.001 get an accuracy range ≥ 96% and on average are able to achieve 100% accuracy from the detection test results of the best value performed for each sample variation of the image object. This research provides significant contributions or benefits in finding the optimal hyperparameter configuration, minimizing overfitting, and shortening the search time for TM accuracy in image classification, particularly in human face recognition. The findings support the development of more efficient and accurate TMs and provide practical guidance for finding better hyperparameter optimization using the grid-search method approach. The results of this study have implications for improving the effectiveness and accuracy of TM models and their development in mobile web applications