PersonaMQA: Context-Aware Persona Inference Engine for Privacy-Preserving Cross-Platform Mobile Messaging Automation

The proliferation of messaging platforms (WhatsApp, Telegram, Discord, WeChat, Messenger, Instagram) has fragmented modern communication. Mobile users struggle to maintain authentic relationships across multiple channels while managing inevitable unavailability. Existing auto-reply solutions present an unresolved tradeoff: cloud systems impose latency (1-3 seconds), recurring costs ($1,095 annually per user), and privacy violations, while template approaches generate uniform, impersonal responses that fail to adapt to relationship types. PersonaMQA is a privacy-first framework that embeds quantized language models (QWen 2.5-7B, Phi-3 Mini) directly on Android devices. The system generates contextually appropriate auto-replies from encrypted local chat histories without cloud dependencies. Core innovations include: (1) entropy-weighted persona extraction, learning contact-specific patterns; (2) hierarchical sensitivity classifier, preventing inappropriate automation in sensitive contexts; (3) quantized edge inference, achieving 82% of full-precision model quality on 2GB RAM devices. Evaluation on the PrivateChat-1.2K dataset (1,850 interactions, 50 contacts) shows 83.2% accuracy on new contacts (vs. 67.4% for MobileGPT) and 99.1% accuracy on learned contacts. PersonaMQA reduces operational costs by 99.7% ($3.65/year vs. $1,095/year), maintains complete data sovereignty, achieves sub-1.2 second latency, and passes FTC 2024 Impersonation Rule compliance [13]. Human evaluation confirms a naturalness rating of 4.3/5.0. PersonaMQA enables users to maintain authentic connections while preserving privacy.