Data-gedreven
CV.

We construct the massive scalar propagator for planar gravitational wave backgrounds propagating on Minkowski space. We represent the propagator in terms of the Bessel's function of suitably deformed nonlocal distance functions, the deformation being caused by gravitational waves. We calculate the propagator both for nonpolarized, for the plus (+) and cross (×) polarized plane waves, as well as for more general planar waves both in D space-time dimensions. The propagator is useful for studying interactions of scalar fields on planar gravitational wave backgrounds in the context of interacting quantum fields, in which renormalization plays a crucial role. As simple applications of the propagator we calculate the one-loop effective action, the scalar mass generated by the scalar self-interaction term and the one-loop energy-momentum tensor, all renormalized by using dimensional regularization and renormalization. While the effective action and scalar mass remain unaffected by gravitational waves, the energy-momentum tensor exhibits flow of energy density in the direction of gravitational waves which grows quadratically with the scalar field mass, gravitational wave amplitude and its frequency.

Money laundering sustains criminal enterprises and poses a global threat to financial stability. In 2024, an estimated $3.1T in laundered money flowed through the global economy. Financial institutions are responsible for detecting and reporting suspicious transactions to Financial Intelligence Units (FIUs), which are governmental bodies tasked with identifying money laundering and terrorist financing. Due to limited and often nonspecific feedback from FIUs, many banks have established internal Anti-Money Laundering (AML) teams to manually assess transactions and label transactions as potentially illicit. This process is resource-intensive, costly, and constrained by limited analyst capacity. As a result, supervised machine learning models are increasingly used to monitor transactions, and are typically trained on labeled data covering less than 2% of all transactions, making the selection of transactions for review a key component for effective money laundering detection.

This study introduces a comprehensive active learning framework designed to detect money laundering from limited labeled data. The framework accounts for the temporal structure of the transactions and dynamically adjusts the classification threshold during training at each iteration by maximizing a net value (NV) objective on a validation set. The net value is defined as: NV = (b - c) · TP - c · FP - (b - c) · FN, where c is the cost of investigating an alert and b represents the benefit of detecting an illicit transaction. Although c is roughly estimated, b is varied to simulate different risk preferences of the financial institution; higher values indicate a risk-averse preference, while lower values reflect a risk-seeking approach. The framework evaluates both supervised and unsupervised query strategies, including a novel explainability-guided method, referred to as SHAP-guided profiling, which prioritizes transactions that are most dissimilar to the average SHAP feature importance profile of legitimate transactions. Throughout the active learning process, we explore how the interaction between the benefit per TP b and the query strategy influence recall, precision, and true negative rate (TNR).

During active learning using only 2.08% of the labeled data on the synthetic dataset AMLworld, a performance evaluated on the validation set is achieved that is comparable to passive learning (model trained on the entire training dataset). After applying feature selection and hyperparameter optimization, active learning continues to perform competitively across strategies, compared to passive learning, where query by committee achieves a higher recall and precision than passive learning. SHAP-guided profiling performs worse than random sampling in both recall and precision. Due to the limited interpretability and poor performance SHAP-guided profiling seems to not be suited for a high-risk domain such as money laundering detection, where performance and accountability are critical.

Comparing different values for b while keeping the query strategy fixed reveals that a larger b increases recall and decreases the TNR. This is in line with the risk preference framework. The results demonstrates the feasibility of assigning an economic value to correctly identify illicit transactions based on the risk preference of the financial institution.

The work concludes with a critical discussion on limitations and domain-specific challenges, such as synthetic datasets, evolving criminal tactics, and severe class imbalance. Several directions for future work are proposed to improve both the methodological approach and practical deployment of the detection model, such as introducing nonlinearity in the net value equation, suggestions for additional explainability-guided query strategies, and varying the number of labeled transactions.

Internal auditors play a key role in ensuring artificial intelligence (AI) compliance with the EU AI Act. This article will examine how Explainable AI (XAI) can play a critical role in assessing AI systems for meeting the specific requirements of transparency, human oversight, and fairness. When effectively implemented, XAI enables traceability, accountability, intervention in AI decisions and can be used as a tool by internal auditors. Effective AI compliance auditing requires understanding of the methods for AI monitoring, associated documentation, and user feedback mechanisms to assess risks, regulatory requirements, and ethical standards.

Law firms operate under stochastic demand while facing capacity constraints, making case acceptance and workload management critical operational decisions. This thesis develops a simulation-based framework to evaluate case acceptance policies using real data from a mid-sized Dutch law firm. Client-level case arrivals are modeled via a Negative Binomial process to capture overdispersion and client heterogeneity, while case workloads and revenues are assigned based on fitted historical distributions. In addition, a billable capacity simulation is constructed based on the historical workforce which takes contractual adjustments into account.

The framework compares acceptance policies under daily and monthly capacity constraints over a multi- year horizon. Results highlight trade-offs between revenue generation and workload stability: unconstrained acceptance maximizes revenue but causes unrealistic workload, whereas capacity-aware policies stabilize workload at the cost of lower revenue. Introducing an allowed monthly overtime improves expected net revenue, but only up to a finite threshold, as exponentially increasing overtime costs eventually outweigh marginal gains. A reliable estimate of the optimal overtime level is determined with a symmetric finite- difference estimator and an asymptotically unbiased confidence interval is provided with the non-overlapping batch means scheme.

Overall, the study demonstrates how simulation-based decision support can quantify long-term trade-offs between revenue and capacity under uncertainty. While abstracting from certain operational complexities, the framework provides a flexible and extensible foundation for evaluating case acceptance policies in legal services and possibly other professional service organizations.