Large Language Models (LLMs) are becoming revolutionary  tool in medical care, but their adoption to clinical practice is challenging by issues of interpretability, dependability, and computational effectiveness. This article introduces a LoRA fine-tuned medical LLM that is intended to adopt parameter-efficient adaptation combined with domain-sensitive reasoning for enhanced clinical decision support. Utilizing benchmark datasets such as PubMedQA and MedMCQA, the base model Apollo-2B is optimized via Low-Rank Adaptation (LoRA) to attain domain alignment without the exorbitant expense of full-scale training. Robust handling of textual as well as numerical queries, including interpretation of lab values led by medical standards, is made possible with a dual-mode reasoning mechanism. In addition to accuracy, the system incorporates structured interpretability via counterfactual explanations, providing clinicians with clear understanding of alternative outcomes and decision routes. This strengthens trustworthiness and use in high-stakes settings including primary care, emergency medicine, and specialist appointments. Deployment through a simplified web interface ensures real-time availability, and the model's output format—Answer, Explanation, and Counterfactuals—is supportive both of clinical teaching and evidence-based practice. With bridging efficiency, interpretability, and explainability, this study fills a pivotal gap in the translation of medical AI from research to practice and responds to the increasingly demanded transparent, adaptive, and clinically informed LLMs.