Senior Financial Crime & Analytics Specialist — FIU investigations · screening & model validation · explainable AI

3+ years across financial crime at Revolut (FIU), PwC (Financial Crime Unit) and AML RightSource: 2,000+ cross-border investigations spanning 12,000+ accounts, SAR/disclosure packages for national FIUs, high-risk CDD/EDD. MSc in Data Science & Society (Tilburg University). I work at the junction the industry struggles to staff: translating financial-crime risk into data, thresholds and governance that survive an auditor.

Independent, fully reproducible validation of the open-source screening stack used in production compliance (OpenSanctions nomenklatura / yente) — run the way a model-risk function would.

• 19.1M live registry records screened (UK PSC + GLEIF vs the EU consolidated sanctions list): 24.5% of review-tier signals are invisible to exact name matching — while exact matching wins on the industry benchmark. The gap is an evaluation artifact, and quantifying it is the core finding.
• Capacity-calibrated thresholds (recall per alert budget, not F1), segment error analysis (cross-script matches missed 10–25× more often), feature-level decision traces, model card and validation report.
• 755k-pair benchmark audit: class balance, label provenance, leakage-safe splits, score-saturation diagnosis.

Can you govern a fraud model whose explanations change between retrains even when performance doesn't?

• 590,540 real e-commerce transactions, strict out-of-time validation, 30-seed stability study across 435 model pairs.
• Top-5 SHAP feature agreement: 0.692 despite statistically indistinguishable PR-AUC — raised to 0.911 via class weighting and seed ensembling.
• Conclusion with governance teeth: explanation stability must be validated and monitored separately from predictive performance.

Python SQL pandas scikit-learn XGBoost SHAP multiprocessing at 19M-record scale Power BI Tableau — applied to transaction monitoring, sanctions screening, fraud typologies, EDD/KYC and model-risk documentation.

LinkedIn · karol.duda113@gmail.com · Polish/EU citizen, unrestricted EU work authorisation