EvalMORAAL
Evaluation framework for cross-cultural moral alignment of large language models. Companion code for the paper published at *SEM 2026.
Projects
Projects
Research code, applied machine learning systems, and earlier engineering work. Where a project produced a paper, a grant or a package, the card links to it.
01Featured
Current & recent
Cultural moral judgments with LLMs
Tests whether large language models capture cultural variation in moral judgments and social norms across societies, toward more culturally aware and fair AI. Published in the Computational Linguistics in the Netherlands Journal (2026).
LLM annotation reliability
Methods for judging when LLM annotations and explanations can be trusted in sensitive domains such as hate speech detection. Paper presented at the GeBNLP workshop, ACL 2025.
CV Priority Sorter
Active-learning and LLM system that ranks incoming CVs for recruiters, reducing hiring time by about 20%. Part of a wider set of generative AI pipelines for insight extraction and de-identification of CVs. Private repository.
ReDOSE
CUCo-funded project on building a circular pharmaceutical industry with AI. A collaboration with Wageningen University & Research, Eindhoven University of Technology, and UMC Utrecht.
FIRMBACKBONE sustainability-analysis CLI
Command-line tool that extracts sustainability keywords and scores company websites on sustainability content, built for the FIRMBACKBONE project.
Explainable sexism identification in social networks
Text-classification models for sexism detection paired with token-level explanations, human evaluation of the generated rationales, and reinforcement-learning-based refinement.
Interpretable financial fraud detection
Combined unsupervised and supervised learning on imbalanced credit-card data, with uncertainty-aware deep learning to stay robust as fraud patterns shift.
02Open source
Evaluation and experiment tooling
Alongside the projects on this page I maintain a set of open-source libraries for auditing the numbers that machine learning runs on: whether an LLM judge can be trusted, whether an A/B readout supports the decision, whether a leaderboard’s order is real, whether a labelled dataset is sound. They live on their own page, with install lines, findings and DOIs.
03Archive
Earlier work (2016–2021)
Machine learning, operations research, simulation, and business intelligence projects from industry and academia.
- Music recommendation with cognitive factors — autoencoder-based recommender that folds mood, context, and human feedback into music suggestions.
- Explainable NLP survey — systematic literature review of explainability methods in NLP and the role of human evaluation.
- Video advertising optimization — testing ad count and placement across user groups at Cafe Bazaar to balance revenue and viewer satisfaction.
- Online retailer customer segmentation — unsupervised mining of purchase data for a Spanish retailer, with a purchase-history recommender.
- Warehouse location by GIS clustering — siting new warehouses across Iran and proposing space, robotics, and same-day-delivery improvements.
- Intelligent bicycle collection — demand, trip, and bike-health forecasting with geographic clustering to prioritize shared-bike pickups.
- Bicycle depot placement — geodata model selecting the best depot per time period from logistics cost, bike counts, and density.
- High-risk rider detection — spatial joins plus classifiers (logistic regression and LDA, about 85% accuracy) to flag bike-share rule violations.
- Bicycle battery health analysis — predicting charge levels, abnormal discharge, and battery failure in shared bikes.
- Preventive bicycle maintenance — failure prediction from distance, time, and trip counts to schedule repairs before breakdowns.
- RFM user segmentation — classifying bike-share users by recency, trip frequency, and average spend.
- Smoking and lung cancer dynamics — system-dynamics model of policy scenarios linking cigarette prices and prevention budgets to smoking rates.
- Hospital patient-flow simulation — discrete-event model of outpatient workflow cutting total length of stay by 9.9% without added resources.
- Energy efficiency in wireless body area networks — priority queueing (G/M/1) model for sensor traffic in WBANs.
- Kidney donation queue optimization — queueing and perishable-inventory model minimizing costs in the organ-transplant supply chain.
- Urban waste management optimization — multi-objective routing model balancing cost, pollution, and biogas production under uncertainty.
04Books
Books
Statistical Reinforcement Learning with Application in Dynamic Pricing Problem
Arshadan, 2022 · Persian
Learning optimal prices from customer interaction with a Bayesian approach: Thompson sampling under strategic and myopic customer behavior, aimed at maximizing long-term revenue.
Machine Learning-Based Fault Diagnosis for Industrial Engineering Systems
Arshadan
Machine learning and deep learning techniques for precision compensation and fault diagnosis in precision motion systems and rotating machinery, with experimental case studies.
Simulation of Residual Stress Measurement Methods in Abaqus Software
Arshadan
Destructive and non-destructive residual stress measurement techniques, with emphasis on nanoindentation and finite-element simulation in Abaqus.