Electronic Health Records (EHR) store valuable patient-staff interaction data. These notes, often unstructured to save healthcare personnel time, can be challenging to analyze manually. Proprietary online LLMs have demonstrated impressive results in analyzing EHR notes. However, Clinical NLP faces unique challenges due to the sensitive and specialized nature of the data. Sending patient information via external APIs poses privacy risks, and hospitals require customized NLP systems to align with their practices. Developing customized LLMs using specific training datasets is crucial to address these challenges. We propose generating synthetic training data using keywords extracted without confidential information. Furthermore, we introduce a reward mechanism that iteratively refines the quality of synthetic documents. This involves scoring synthetic candidates against real clinical reports using a semantic textual similarity score and performing an alignment step to align the model with its best-scored utterances.