Despite the ideal of a unified Single Market, a powerful “home bias” pervades EU public procurement, hinting at unseen barriers that conventional analysis fails to capture. This study introduces an interpretable AI framework to investigate these dynamics, pairing a LightGBM model with SHapley Additive exPlanations (SHAP) to examine the vast Tenders Electronic Daily (TED) database (2018–2023). Concretely, we propose a fuzzy linguistic layer that translates SHAP’s complex quantitative outputs into intuitive, human-readable terms. Our model effectively distinguishes local from non-local awards (AUC ≈ 0.855), revealing that while high-value contracts expectedly attract broader competition, the most potent predictors are a country’s own history of local awards and structural factors like the buyer’s type and location. This points not to isolated incidents, but, rather, to deep-seated patterns shaping market fairness. Our combined XAI-Fuzzy approach offers a new instrument for transparent governance, enabling policymakers to diagnose market realities and forge a more genuinely open and equitable European public square.

Key field information extraction from documents is an increasingly covetable task. Previous related work has touched upon the subject through the lens of rule-based systems or through natural language processing methods. In this paper we see the task of information extraction from invoices as an object detection task. To this end, we used three different models YOLOv5, Scaled YOLOv4 and Faster R-CNN to detect key field information in invoices. Additionally, we propose a data preprocessing method that helps to better generalize the learning. All of the experiments were performed on a custom made dataset with a very high variety of invoice layouts. This decision comes in part from the lack of any suitable public dataset and from the need of finding the best procedure for annotating data pertaining to this task. The obtained results were encouraging, leading us to the conclusion that object detection is a viable method for information extraction.

Automatic car damage detection has been a topic of significant interest for the auto insurance industry as it promises faster, accurate, and cost-effective damage assessments. However, few works have gone beyond 2D image analysis to leverage 3D reconstruction methods, which have the potential to provide a more comprehensive and geometrically accurate representation of the damage. Moreover, recent methods employing 3D representations for novel view synthesis, particularly 3D Gaussian Splatting (3D-GS), have demonstrated the ability to generate accurate and coherent 3D reconstructions from a limited number of views. In this work we introduce an automatic car damage detection pipeline that performs 3D damage segmentation by up-lifting 2D masks. Additionally, we propose a simple yet effective learning-free approach for single-view 3D-GS segmentation. Specifically, Gaussians are projected onto the image plane using camera parameters obtained via Structure from Motion (SfM). They are then filtered through an algorithm that utilizes Z-buffering along with a normal distribution model of depth and opacities. Through experiments we found that this method is particularly effective for challenging scenarios like car damage detection, where target objects (e.g., scratches, small dents) may only be clearly visible in a single view, making multi-view consistency approaches impractical or impossible. The code is publicly available at: https://github.com/DragosChileban/CrashSplat.

Vision-Language Models (VLMs) trained on large-scale multimodal data develop latent object detection capabilities through grounding pretraining, yet these remain largely untapped for generalpurpose detection. We systematically investigate extracting detection from VLMs without task-specific training or architectural modification, a setting we term category-prompted zero-shot detection, where a target vocabulary is provided at inference time but no detection-specific supervision is used. Our central finding is that VLMs represent bounding boxes in a normalized 1000 × 1000 coordinate space, a convention established by Qwen-VL and adopted by multiple model families, that differs from pixel coordinates. Simply rescaling outputs from this internal space to image dimensions yields up to 48% F1 improvement. Prompt engineering, by contrast, shows negligible effect (±1–4% F1) regardless of how explicitly coordinate systems are specified, consistent with this representation being learned during pretraining. The best-performing model achieves 38.2% mAP@50 on COCO, reaching 61% of supervised DETR performance, with consistent results across three benchmarks spanning photographic and non-photographic domains. Strategies that reliably improve CNN-based detectors, including image tiling, model cascading, and ensembling, consistently degrade VLM detection: tiling destroys global scene context, while cascading and ensembling fail because VLMs detect complementary rather than redundant object sets. Small objects remain a fundamental limitation, with recall below 10% for objects under 32×32 pixels.We provide practical guidelines for practitioners, showing that post-processing choices, particularly coordinate rescaling and label normalization, matter far more than prompt optimization.