Industry — AcademicTransfer. I lead AI and data-science work across CV–vacancy matching and ranking, LLM content optimisation, recruiter analytics, and end-to-end ML tooling for the two-sided Dutch academic-jobs marketplace (22 research universities and university medical centres).

Research — Utrecht University. My doctoral thesis develops explainable NLP across the full LLM life cycle, from token-level SHAP analysis to cross-cultural moral-alignment evaluation of LLMs. Recent work centers on LLM evaluation: LLM-as-judge frameworks (EvalMORAAL) and preference optimization under annotator disagreement (RA-DPO, MAP-PO).

I work where engineering rigor meets explainability research — shipping models that deliver in production and expose why they make each decision.

Utrecht University, 2026

A six-chapter empirical thesis on explainability across the full LLM life cycle: a survey of XAI for NLP, a transparent BERT pipeline for online sexism detection, SHAP-driven probing of AI-text-detector robustness, content-vs-demographic explanations for LLM annotators, cross-cultural moral-alignment evaluation of 26 LLMs against the World Values Survey and PEW, and the EvalMORAAL chain-of-thought-plus-LLM-as-judge framework benchmarking 20 LLMs across 64 countries.

Each chapter has a paper and a public companion repository with citation metadata and a tagged release.

Full publication list on mohammadi.cv and Google Scholar.

LLM and ML systems serving the Dutch academic job market, end to end: data pipelines, model serving, monitoring, and recruiter-facing tools.

• Semantic CV-to-vacancy matching and LLM-assisted priority ranking of applicants, evaluated with A/B tests, uplift analysis, and bandit simulations
• LLM-based job-description optimisation and vacancy-text rewriting (OpenAI API)
• Recruiter analytics and model-monitoring dashboards
• CRM and workflow automation for recruitment teams

Internal repositories at @academictransfer (private; access on request):

As Head of Data Science & BI I built the operations-intelligence layer for a city-scale bike-sharing fleet: bicycle-transportation-intelligence — a Streamlit dashboard on live Oracle fleet data, with GeoPandas geofencing and H3 spatial indexing plus folium maps to plan rebalancing and collection.

Earlier: Senior Data Scientist at SnowaTec (2021–2023) — details on mohammadi.cv.

The evaluation libraries are one line of work rather than a collection. Each takes a single way evaluation goes wrong — a judge with a position bias, a label set nobody checked, a leaderboard read past its own confidence intervals, an experiment called too early — and makes it measurable, with the headline finding stated as a number instead of a claim. evalstack is the map of all eleven: what each measures, what it found, and the two chains that run end to end.

• EvalMORAAL — chain-of-thought plus LLM-as-judge evaluation of moral alignment, benchmarking 20 LLMs across 64 countries against the World Values Survey and PEW (*SEM 2026).
• trajectory-judge — how much an LLM judge misses when an agent reaches the right answer the wrong way: a tool-using environment with injected faults, five judge designs compared on detection, step localisation, calibration and cost.
• judgekit — bias probes for LLM judges, each with a bootstrap confidence interval: position, verbosity, self-preference and calibration, where qwen2.5:14b picks whichever answer is shown first 77% of the time.
• judgepanel — estimate how accurate your judges are without any gold labels: Dawid–Skene EM over the disagreement pattern alone, which exposes a flag-everything judge exactly (estimated specificity 0.000).
• abkit — sample-ratio mismatch, peeking, multiple testing and winner's-curse checks for experiment readouts, applied to 4,873 real Upworthy tests where 16% fail the sample-ratio check.

• spark-search-ranking — counterfactual learning-to-rank for marketplace search logs in PySpark: position-bias estimation, IPS-weighted training, NDCG evaluation with a full test suite.
• rankkit — ranking metrics with error bars, and click metrics that survive position bias: on the bundled example, counting raw clicks rejects a ranker that is genuinely better and correcting for position reverses the verdict.
• dynamic-pricing-dashboard — dynamic-pricing simulator with Thompson sampling demand learning and forward-looking customers, running fully in the browser via WebAssembly (live demo).

• awesome-explainable-nlp — 145 papers, tools, datasets, tutorials and venues on explainability for NLP and LLMs, with weekly automated link checking. Contributions welcome.

The other six evaluation libraries — raterkit, retainkit, calikit, abeval, arenakit and judgewatch — are described in evalstack.

Also maintained: MAP-PO and RA-DPO (preference optimization under annotator disagreement), explainkit (explainability audits), ml-foundations · modern-ai-engineering (from-scratch ML and production LLM notes), and guides to ML learning paths, European summer schools and Dutch AI master's programmes.

Elsewhere: FBB Sustainability Analysis — an environmental-impact analysis CLI on a Dutch firm panel, built for the Firmbackbone research infrastructure.

Everything else is in the repositories tab.

Open to applied AI roles and consulting in NL / EU, and to research collaboration on explainability, LLM evaluation, and cultural alignment.

mohammadi.cv · LinkedIn · ORCID · [email protected]