<p>Large language models (LLMs) are increasingly used in AI applications, making it important to address biases, particularly gender bias, to reduce stereotypes and promote fairness. Biased LLMs can misrepresent information, potentially contributing to social inequality and reduced trust. This research focuses on detecting and mitigating bias, with particular attention to gender bias. Specialized metrics—Disparity Index, Idea Consistency Score, Thematic Consistency Score, and Zero-Shot Classification were used to evaluate model behavior across sensitive factors in Hindi and English prompts. These metrics were applied to analyze responses to diverse prompts in both languages, helping detect disparities in model outputs across sensitive factors such as gender. Based on insights from these evaluations, two approaches were developed to address these biases. Prompt engineering was first applied to refine model outputs and reduce bias. Building on these improvements, bias reduction was further achieved through Low-Rank Adaptation (LoRA)-based fine-tuning, a resource-efficient technique that achieved further bias reduction. Initial prompt engineering showed a reduction in polarized responses by approximately 40% and improved positive portrayals by 45%. Fine-tuning with LoRA achieved a 39.6% reduction in gender bias, 27.8% reduction in racial bias, and 10.6% reduction in socioeconomic bias, along with a 19.7% increase in positive portrayals, demonstrating effective and resource-efficient bias mitigation.</p>