EchoTwin Privacy-Aware Personalized Email Reply Generation with Low Rank Adaptation

Abstract

Email continues to serve as a primary medium for both professional and personal communication. Although contemporary assistance systems, including Smart Reply and Smart Compose, enhance writing efficiency, they generally do not explicitly model or adapt to an individual’s unique communication patterns. This paper presents EchoTwin, an AI-assisted email-reply generation system that produces personalized draft replies aligned with a user’s historical communication patterns. EchoTwin builds on a decoder-only Transformer large language model and applies Low-Rank Adaptation (LoRA) for parameter-efficient personalization: a lightweight adapter is trained for each user while a single base model is shared across the population, avoiding the cost of an independent model per user. The system integrates with Gmail through Google OAuth 2.0 and the Gmail API, keeping a human reviewer in the loop to approve, edit, or reject every draft before it is sent. The approach was evaluated on a curated subset of the Enron Email-Reply dataset, comprising 2,221 email-reply pairs from ten users, comparing personalized generation against the base model on 223 held-out samples using lexical similarity (ROUGE), semantic similarity, corpus-level style similarity, and stylometric analysis. The personalized adapters achieved statistically significant gains of 8.10% in ROUGE-1, 35.65% in ROUGE-L, and 5.64% in corpus-level style similarity, plus a 25.26% reduction in mean stylometric distance (Wilcoxon signed-rank test, ???? < 0.05). Sentence-BERT similarity to the reference reply decreased by 10.29%, reflecting the expected trade-off between reference-based similarity and individual style adaptation. These prototype-level results indicate that user-specific LoRA adapters can improve measured stylistic alignment in personalized email-reply generation while keeping the user in control of the final output.