PhD Position Decentralized and Trustworthy AI Pipelines (EU project WALTZ)

Public administrations hold the data for better public services but cannot pool it. We aim to build decentralized AI pipelines that let them learn from data they can never expose, in the EU project WALTZ.

Job description

The Horizon Europe Innovation Action WALTZ (Workflow-Driven AI-Enabled Lawful and Trusted Data Spaces for Public Authorities) builds a trusted "system-of-systems" data stack that turns heterogeneous administrative data into governed, AI-ready assets, validated in six public-sector pilots. TU Delft leads its task on AI-ready training and evaluation pipelines using real and synthetic data. Public authorities cannot ship their data to a central trainer, the participants in such a pipeline cannot all be assumed honest, and the models they increasingly want to deploy are LLM-based, whose failures are semantic rather than crashes.

We aim to build training and inference pipelines that work across administrative boundaries, and evaluation methods that say something honest about how far they can be trusted.

Initial concrete task ideas include:

• Design decentralized and federated pipelines that combine real and synthetic data across organisational boundaries without a central aggregator, under controlled mixing strategies that address data scarcity and distribution shift.
• Make them resilient to Byzantine participants, and to poisoning of the synthetic data supply.
• Quantify what leaks: membership inference and reconstruction against models trained on real-synthetic mixes, and the utility cost of differentially private generation.
• Treat LLM-based components as a distributed system: replicate inference across diverse models, study how their failures correlate, and use quorum and semantic-agreement mechanisms to turn individually unreliable answers into trustworthy ones.
• Build the evaluation side: benchmarks, cross-validation schemes, reporting protocols — covering accuracy, robustness, bias, generalisation and privacy leakage, released as open, reproducible software.

You will be based at the Faculty of Electrical Engineering, Mathematics and Computer Science (EEMCS), in the Distributed Systems group of the Department of Software Technology, supervised by Dr Jérémie Decouchant, whose work spans Byzantine fault tolerance, decentralized learning and trustworthy AI systems. TU Delft's DelftBlue and DAIC GPU clusters are available for large-scale experiments.

WALTZ has more than thirty partners, so your work will not stay in the lab: you will collaborate with research groups across Europe and see your methods stress-tested by real public authorities. Alongside your research you will publish at top venues, follow the TU Delft Graduate School doctoral programme, and take on a modest teaching and supervision load (up to 15%).

Job requirements

• An MSc (completed, or close to completion) in Computer Science, Data Science, Artificial Intelligence, Electrical Engineering or a closely related field.
• A solid foundation in machine learning and/or distributed systems. Familiarity with federated or decentralized learning, fault tolerance, generative models, privacy-preserving ML, or LLM-based systems is a strong asset.
• Strong programming skills in Python and hands-on experience with a modern deep learning framework (e.g. PyTorch), including work on Linux-based GPU/HPC clusters.
• Demonstrable interest in trustworthy AI: privacy, robustness, adversarial behaviour, evaluation methodology and reproducibility.
• Strong analytical skills and the independence to carry a research agenda over four years, combined with the discipline to meet project deliverable deadlines.
• Excellent command of written and spoken English. Non-native speakers without an English-taught degree must meet the TU Delft English language requirements (e.g. TOEFL iBT 90 or IELTS 6.5 overall).
• A collaborative attitude: you enjoy working in a large international consortium and can explain your work to non-academic stakeholders such as public administrations.
• Commitment to open science: publishing code and benchmarks, and contributing to open-source software.

TU Delft (Delft University of Technology)

TU Delft is a top international university combining science, engineering and design. It delivers results in education, research and innovation to address challenges in energy, climate, mobility, health and digital society.

Faculty of Electrical Engineering, Mathematics and Computer Science

The Faculty of Electrical Engineering, Mathematics and Computer Science (EEMCS) brings together three scientific disciplines that reinforce each other and drive the technology used in daily life, including AI. The faculty offers strong international research, innovative engineering education, and excellent labs and facilities. You will work within TU Delft's Faculty of Electrical Engineering, Mathematics and Computer Science, where research in software technologies, AI, and applied mathematics supports ground-breaking work and strong international collaboration.

Conditions of employment

Doctoral candidates will be offered a 4-year period of employment in principle, but in the form of 2 employment contracts. An initial 1.5 year contract with an official go/no go progress assessment within 15 months, followed by an additional contract for the remaining 2.5 years assuming everything goes well and performance requirements are met.

As a PhD candidate you will be enrolled in the TU Delft Graduate School. The TU Delft Graduate School provides an inspiring research environment with an excellent team of supervisors, academic staff and a mentor. The Doctoral Education Programme is aimed at developing your transferable, discipline-related and research skills.

The TU Delft offers a customisable compensation package, discounts on health insurance, and a monthly work costs contribution. Flexible work schedules can be arranged.

As part of knowledge security, TU Delft conducts a risk assessment during the recruitment of personnel. The assessment is based on information provided by candidates and takes place at the final stages of the selection process. When the outcome of the assessment is negative, the candidate will be informed.

#EUfunded This is an EU funded project, named WALTZ, within program Horizon Europe.