PersonalAI · PersonalAI: Efficient and Private Text Personalization with Large Language Models

The rapid adoption of Large Language Models (LLMs) for text generation has created a critical tension between functionality and privacy: while users increasingly rely on services like ChatGPT for personal and business communications, such cloud-based systems require surrendering control of sensitive data to third-party companies, often outside EU jurisdiction. The PersonalAI project addresses this fundamental challenge by enabling fully private, on-device personalization of open-source LLMs, leveraging breakthrough compression and fine-tuning techniques developed in the PI's ERC Starting Grant ScaleML project. Our preliminary PanzaMail study showed that models fine-tuned on as few as 50-100 user emails can generate text closely following genuine user writing, while our novel RoSA method enables this personalization on commodity hardware—reducing computational requirements by over 90% compared to standard approaches.The technical innovation of the PersonalAI proof-of-concept lies in a four-component system that triages queries by complexity, executes simple tasks locally with efficiently-trained personalized models, sanitizes sensitive information from complex queries before cloud processing, and re-personalizes external outputs to match user style. As such, PersonalAI targets the rapidly-expanding EU AI market, where data sovereignty and GDPR compliance are paramount, for both companies and individual users. It offers the first solution that combines the personalization quality of proprietary services with complete data privacy. The project assembles a world-class team including the PI, lead developer Eugenia Iofinova (architect of PanzaMail), and two MIT advisors with deep entrepreneurial expertise. By transforming academic breakthroughs into a commercially viable product, PersonalAI will establish a new paradigm where users maintain complete ownership of their data and AI models while accessing state-of-the-art text generation and model editing capabilities.