BilAI Key Researcher Dan Alistarh receives ERC Grant for Private AI on Your Device

25.03.2026

BilAI key researcher Dan Alistarh has been awarded his second European Research Council (ERC) Proof of Concept Grant, for his project PersonalAI,which aims to bring fully private, personalised text generation to everyday devices.

Large Language Models such as ChatGPT, Gemini, and Claude have become indispensable tools for millions of users. Yet, using them typically means sending sensitive data to remote cloud servers. Together with his team, Dan Alistarh, Professor at the Institute of Science and Technology Austria (ISTA) and Key Researcher in the Bilateral AI Cluster of Excellence, is determined to change that.
His newly funded project, PersonalAI, develops a framework that enables users to personalise open-source language models directly on their own devices, such as laptops, PCs, or even phones, without sharing any personal data with third parties. This grant provides €150,000 over 18 months to bring this research closer to a real-world product.

From efficient ML research to a privacy-first product

The project builds directly on breakthroughs from Alistarh's ERC Starting Grant ScaleML, including the widely adopted GPTQ model compression method and the RoSA parameter-efficient fine-tuning technique. These innovations, which also connect to his research within BilAI on efficient machine learning, make it possible to train and run AI models on commodity hardware — reducing computational requirements by over 90% compared to standard approaches.
A preliminary study called PanzaMail, designed and led by PhD researcher Eugenia Iofinova, already demonstrated that language models fine-tuned on as few as 50–100 user emails can convincingly replicate a person's writing style. In user studies, contacts of the test users perceived the AI-generated emails as genuinely written by that person, not by another human or a machine. Crucially, the entire training and inference pipeline ran on a standard laptop, showing that on-device personalisation is not only possible but practical.

A four-step system for privacy and quality

PersonalAI's architecture addresses the fundamental challenge that small, on-device models cannot yet match the capabilities of large cloud-based systems for every task. The solution is to triage incoming queries by complexity, handles simple tasks locally with a personalised model, sanitises sensitive information from complex queries before sending them to a larger cloud model, and re-personalises the returned output to match the user's style. This approach combines the quality of leading commercial services with full data privacy, which is particularly relevant in the European market.

Strong connections to BilAI

PersonalAI connects directly to BilAI's GreenAI research module, which develops efficient methods for LLM training and inference. The compression and fine-tuning techniques built in GreenAI at the heart of PersonalAI, enabling models to run and be trained on everyday hardware, and are a prime example of how BilAI research can unlock entirely new capabilities, in this case enabling full data privacy by making on-device personalisation practical. The project demonstrates how fundamental research within the Cluster can translate into technology with real societal impact, empowering users to benefit from state-of-the-art AI while retaining full control over their data.

The team behind PersonalAI

A central role in the project is played by Eugenia Iofinova, a PhD researcher in Alistarh's group at ISTA and the architect of PanzaMail. Iofinova designed the training and evaluation protocols that demonstrated on-device personalisation is feasible, and she will lead the technical development of the PersonalAI product. The team also includes systems engineer Erik Schultheis,  XISTA innovation manager Edmundo R. Sánchez Guajardo, and scientific advisors from MIT.