Accurate extraction of pharmacologic information from unstructured clinical notes is essential for downstream applications in healthcare informatics. The novelty of this work lies in a hybrid medication extraction pipeline that integrates broad coverage Natural Language Processing (NLP) with Large Language Model (LLM). Specifically, we employ QuickUMLS and ScispaCy (UMLS entity linker) to achieve high recall mapping of all drug mentions across 500 de-identified MIMIC-IV patient records, followed by Llama 3.3, which refines and filters these candidates to retain only the medications actually administered. The approach is evaluated against a curated ground truth of administered drugs, reporting precision, recall, and F1 score. QuickUMLS attains broad coverage but lower F1 scores due to ground truth constraints, whereas ScispaCy provides more conservative, model-based linking. Llama 3.3 demonstrates consistently high precision and recall, even under API batch limitations (30–40 patients). Results remain stable across JSON, CSV, and XLSX output formats. Despite limitations stemming from the modest cohort size and QuickUMLS’s over inclusion tendency, findings indicate that the NLP–LLM synergy provides an effective and scalable solution for clinical medication extraction. Future work will extend validation to diverse EHR systems, larger datasets, and multilingual environments.