Jarosław Kurek

Large online marketplaces increasingly rely on multi-stage ranking pipelines where a learned relevance model is complemented by business-aware constraints such as contractual pacing, exposure caps and commercial alignment objectives. This paper develops a second-stage, contract-aware re-ranking layer for real-estate recommendation that explicitly balances user–item relevance with plan fulfillment, lead value and operational guardrails. The proposed multi-objective re-ranker (PMOR) combines a calibrated base relevance score with multiplicative business adjustments and subtractive penalties for approaching contractual caps and for within-slate similarity. The method supports heterogeneous settlement models, including pay-per-action and fixed-fee contracts, via contract-specific weights. Because the scoring function is deterministic and structured, it admits exact component-wise contribution analysis and counterfactual ablations without relying on surrogate explainability methods. Offline evaluation on production logs from an anonymized marketplace covers 4219 recommendation requests and 184,147 candidate items, joined with daily business snapshots using an as-of strategy to prevent look-ahead bias. Under a profit proxy based on effective lead value, position discounting and billability, PMOR achieves an indexed expected-revenue proxy of 487.4 (baseline = 100), corresponding to a lift of 387.4% over a model-only baseline and 48.4% over a legacy production re-ranker (LPR). The gain is primarily associated with improved billable exposure, increasing the share of billable positions in TOP-3 to 78.22% compared with 37.85% for LPR. We discuss parameter sensitivity, operational considerations and limitations of offline proxy objectives for deployment.

Background: Accurate vertebral instance segmentation on full-spine radiographs is essential for spinal parameter assessment, but supervised methods require costly instance-level annotations and may be sensitive to domain shift. Methods: We investigated whether promptable segmentation foundation models can generalize zero-shot to raw DICOM spine radiographs without task-specific training. We evaluated SAM-ViT-Huge, SAM2-Hiera-Large, and MedSAM-ViT-Base on 144 full-spine radiographs with 1309 annotated vertebral masks using a standardized pipeline for DICOM decoding, intensity normalization, automatic prompt generation, and instance-level evaluation. For each prompt, models produced three candidate masks. Performance was reported under an oracle protocol selecting the candidate with the highest IoU against ground truth and a model-score protocol selecting the candidate with the highest predicted IoU. Metrics included IoU, Dice, precision, recall, ASSD, and HD95. Results: The best configuration was SAM-ViT-Huge with rectangle prompting, reaching a mean IoU/Dice of 0.782/0.870 under oracle selection and 0.737/0.837 under model-score selection. SAM2-Hiera-Large with rectangle prompting followed (0.744/0.848 oracle; 0.699/0.815 model-score), ahead of MedSAM-ViT-Base (0.599/0.737 oracle; 0.387/0.499 model-score). Point prompting yielded consistently low overlap (IoU 0.224–0.319; Dice 0.276–0.414) despite high recall, indicating systematic over-segmentation and large boundary errors. Conclusions: Zero-shot vertebral instance segmentation on raw DICOM spine radiographs is feasible with promptable foundation models when prompts sufficiently constrain target extent. Rectangle prompting is clearly more effective than point prompting in this setting.

Recruitment automation increasingly relies on Large Language Models (LLMs) for extracting structured information from unstructured CVs and job postings. However, production data often arrive as heterogeneous, privacy-sensitive PDFs, limiting reproducibility and compliance. This study introduces a deterministic, GDPR-aligned pipeline that converts recruitment documents into structured, anonymized Markdown and subsequently into validated JSON ready for downstream AI processing. The workflow combines the Docling PDF-to-Markdown converter with a two-pass anonymization protocol and evaluates three LLM back-ends—GPT-4o (Azure, frozen proprietary), GPT-OSS-120B and Llama-3.1-8B-Instruct—using identical prompts and schema constraints under near-zero-temperature decoding. Each model’s output was assessed across 2280 multilingual CVs using two complementary metrics: reference-based completeness and content similarity. The proprietary GPT-4o achieved perfect schema coverage and served as the reproducibility baseline, while the open-weight models reached 73–79% completeness and 59–72% content similarity depending on section complexity. Llama-3.1-8B-Instruct performed strongly on standardized sections such as contact and legal, whereas GPT-OSS-120B better-handled less frequent narrative fields. The results demonstrate that fully deterministic, auditable document extraction is achievable with both proprietary and open LLMs when guided by strong schema validation and anonymization. The proposed pipeline bridges the gap between document ingestion and reliable, bias-aware data preparation for AI-driven recruitment systems.

Understanding and monitoring the phenological phases of trees is essential for ecological research and climate change studies. In this work, we present a comprehensive evaluation of state-of-the-art convolutional neural networks (CNNs) and transformer architectures for the automated classification of the flowering phase of Tilia cordata Mill. (small-leaved lime) based on a large set of real-world images acquired under natural field conditions. The study introduces a novel, automated image quality filtering approach using an XGBoost classifier trained on diverse exposure and sharpness features to ensure robust input data for subsequent deep learning models. Seven modern neural network architectures, including VGG16, ResNet50, EfficientNetB3, MobileNetV3 Large, ConvNeXt Tiny, Vision Transformer (ViT-B/16), and Swin Transformer Tiny, were fine-tuned and evaluated under a rigorous cross-validation protocol. All models achieved excellent performance, with cross-validated F1-scores exceeding 0.97 and balanced accuracy up to 0.993. The best results were obtained for ResNet50 and ConvNeXt Tiny (F1-score: 0.9879 ± 0.0077 and 0.9860 ± 0.0073, balanced accuracy: 0.9922 ± 0.0054 and 0.9927 ± 0.0042, respectively), indicating outstanding sensitivity and specificity for both flowering and non-flowering classes. Classical CNNs (VGG16, ResNet50, and ConvNeXt Tiny) demonstrated slightly superior robustness compared to transformer-based models, though all architectures maintained high generalization and minimal variance across folds. The integrated quality assessment and classification pipeline enables scalable, high-throughput monitoring of flowering phases in natural environments. The proposed methodology is adaptable to other plant species and locations, supporting future ecological monitoring and climate studies. Our key contributions are as follows: (i) introducing an automated exposure-quality filtering stage for field imagery; (ii) publishing a curated, season-long dataset of Tilia cordata images; and (iii) providing the first systematic cross-validated benchmark that contrasts classical CNNs with transformer architectures for phenological phase recognition.

Job–candidate matching is pivotal in recruitment, yet traditional manual or keyword-based methods can be laborious and prone to missing qualified candidates. In this study, we introduce the first Siamese framework that systematically contrasts GRU, LSTM, and Transformer sequential heads on top of a multilingual Sentence Transformer backbone, which is trained end-to-end with triplet loss on real-world recruitment data. This combination captures both long-range dependencies across document segments and global semantics, representing a substantial advance over approaches that rely solely on static embeddings. We compare the three heads using ranking metrics such as Top-K accuracy and Mean Reciprocal Rank (MRR). The Transformer-based model yields the best overall performance, with an MRR of 0.979 and a Top-100 accuracy of 87.20% on the test set. Visualization of learned embeddings (t-SNE) shows that self-attention more effectively clusters matching texts and separates them from irrelevant ones. These findings underscore the potential of combining multilingual base embeddings with specialized sequential layers to reduce manual screening efforts and improve recruitment efficiency.