Federated Learning (FL) enhances data privacy by training machine learning models across devices without centralizing sensitive data, aligning with GDPR mandates. However, FL faces challenges with Non-Independent and Identically Distributed (Non-IID) data, which affects model performance. To address this, we leverage Service-Oriented Architecture (SOA) principles, specifically the Aggregator pattern, by introducing in-client clustering during FL’s local training phase to boost accuracy. This approach is applied to a Named Entity Recognition (NER) task in the medical domain using ADE Corpus and CADEC datasets, with further evaluation on the general-purpose CoNLL dataset for generalizability. Results show improvements in weighted F1-scores: 3.5% for ADE Corpus, 1.4% for CADEC and a marginal gain for CoNLL, highlighting SOA’s potential in optimizing FL. These findings encourage future exploration of SOA principles in FL, offering promising solutions for distributed learning challenges.