The Louvain community detection and other graph-based clustering techniques are appealing to high-dimensional data since they have the ability to infer complicated structural relationships using similarity graphs. Nevertheless, their results are very sensitive to the choice of graph construction and the modularity resolution parameter which can give unstable partitions or under-clustering in the case of manually set hyperparameters. To overcome this shortcoming, this paper suggests a hybrid clustering model that combines the Louvain-based graph clustering with Differential Evolution (DE) to jointly optimize the hyperparameters. The framework proposed will optimize neighborhood size, embedding dimension, pruning quantile, and resolution simultaneously by using a single objective function that balances the modularity, silhouette score, and cluster balance, and a soft cluster-count regularization term is employed to encourage meaningful cluster granularity without enforcing a strict constraint on the number of clusters. In practice, to achieve the DE optimization step, it is done once offline on a representative subsample, with final clustering applied to the entire dataset using the optimized parameters. The results of experiments on five benchmark datasets (MFEAT, COIL-20, HAR, MNIST, and ISOLET) demonstrate that the proposed framework is stable in terms of outperforming the traditional clustering approaches and a fixed-parameter graph baseline. To illustrate, in MFEAT ARI increased by 0.52 to 0.65 and ACC increased by 0.54 to 0.67 whereas in ISOLET the detected cluster count increased by 10.0 to 24.7, which was much closer to the reference value of 26. The findings indicate a practical utility of the framework in high-dimensional clustering and the methodological importance of integrating joint graph-hyperparameter optimization with soft cluster-count regularization to achieve better clustering quality and control cluster-granularity.