Description
Temporary contract | 14 ( months) | BelvalAre you passionate about research? So are we! Come and join us
SUMMARY
As part of Luxembourg’s AI Factory initiative , a fully funded PhD position is open to design and implement a formal and computational framework for structuring the inputs and outputs of AI regulatory sandboxes under the EU AI Act. The position focuses on formalizing legal, technical, and organizational information through mathematical modeling, and on structuring sandbox results into an evolving, linked knowledge base for reuse and transparency.
This is an entry-level scientist working under close supervision. The candidate will perform the job according to plans and directions set up with the thesis supervisor in alignment with the . Organizes own day-to-day activities. Can explain theories, facts, and practices related to own work. Develops knowledge of discipline, RDI methodologies, technical instruments, and protocols. Gains knowledge of LIST RDI activity, it’s partnership ecosystem and the how it responds to market needs.
GENERAL ROLES AND RESPONSIBILITIES
The primarily responsibility of a PhD students is to define their research project and successfully complete it, achieving agreed working plan and objectives in due time. To do so, and depending on their area of activity, their duties may include :
SPECIFIC ROLES AND RESPONSIBILITIES RELATED TO THE PHD TOPIC
The successful candidate will develop the foundations and tools to support three interconnected pillars of regulatory sandboxing under the EU AI Act : (i) formalization of inputs, (ii) development of assessment tools for trustworthy AI, (iii) structuring outputs into a Knowledge base
1. Formalization of Inputs
Model unstructured input information—including legal texts, application forms, onboarding narratives, and risk documentation—using rigorous mathematical formalisms such as graph-based structures, domain-specific languages, or state machines.
Translate these models into actionable orchestrations of AI agents or tools capable of parsing, extracting, and transforming these heterogeneous inputs into a computationally operable format. This enables traceability, semantic alignment, and automated reasoning , supported by embedded quality controls.
The objective is to ensure that the onboarding and compliance preparation phases of sandboxing are grounded in a mathematically precise and interoperable input framework .
2. Development of Assessment Tools for Trustworthy AI
Formalise, from first principles, the core requirements of Trustworthy AI —as defined in the EU AI Act and related ethical and technical literature—into testable and computable properties. These include :
Fairness : Detection of bias, inequality in treatment, or systemic discrimination across demographic groups.
Agentic Behaviour : Evaluation of alignment with human goals, autonomy in task execution, and ethical safeguards in decision-making.
Robustness : Assessment of the AI system’s resilience under input variation, adversarial conditions, or data drift.
Based on these formal representations, develop a library of automated, reusable assessment tools capable of running during sandbox engagements. These tools will allow for technical evaluation, structured feedback, and the monitoring of improvements across iterations.
This component directly supports the core testing phase of regulatory sandboxing, bridging the gap between abstract legal requirements and concrete technical audits.
3. Structuring Outputs into a Knowledge Base
Translate sandbox results—including evaluator feedback, test outcomes, compliance justifications, and final reports—into structured, ontologically grounded representations.
Contribute to building a linked knowledge base that captures sandbox experiences over time, enabling benchmarking, reuse of testing artifacts, and horizontal policy learning.
Design mechanisms for iterative refinement , where outputs from past engagements inform both future assessments and the evolution of the modeling language itself.
The final outcome is a dynamic, queryable infrastructure that serves regulators, auditors, and developers—helping them navigate, compare, and analyse diverse sandbox cases through shared vocabularies and structured evidence.
This knowledge base will serve as a living resource for regulators, developers, and auditors, enabling navigation and analysis across multiple sandbox cases through shared vocabularies and structured evidence.
Profile
REQUIRED QUALIFICATIONS
Starting date
Dès que possible
And • Esch sur Alzette, Luxembourg