Authors - Shripada Rao, Aadithya Mahesh, Navya Jaideep, Rajeshwari Hegde, Vinay Rao, Saurabh Suman Choudhuri Abstract - This paper introduces a novel approach to integrate LLM capabilities directly on mobile devices to enhance chat applications. By implementing a tonality-driven paraphrasing feature, our system can rephrase poorly written messages into clear, professional text while preserving the intended tone. Unlike conventional server-side AI solutions that raise privacy concerns, our approach processes data locally using fine-tuned models (TinyLlama Instruct 1.1B and Qwen2 0.5B) with parameter-efficient techniques such as LoRA and QLoRA. Experimental evaluations demonstrate competitive paraphrasing quality, improved inference speed, and reduced resource consumption on mobile devices, making this work a promising step toward privacy-preserving on-device conversational assistance.
