Daniel Wlazło¶

Data Science Manager at Allegro Pay — trustworthy machine learning for credit risk, in Warsaw.

About¶

I’m a Data Science Manager in credit risk at Allegro Pay, where I lead modelling work for one of Central Europe’s largest consumer‑finance platforms. My focus is application scoring — PD models for individual borrowers — and, increasingly, what trustworthy machine learning looks like in a regulated domain: fairness, calibration, explanation.

I’ve spent close to a decade in credit risk and data science, across Allegro Pay, Hexaware (for AXA), PKO Bank Polski, and ING Bank Śląski. Before that I was a robotics engineer at Samsung and a UAV constructor at Aviation Technik. Robotics was where I started — it’s still how I think about machine learning: a system you can describe, debug, and put under load. Outside work I learn Czech, follow Formula 1, and tinker with robotics.

Notes¶

• May 2026 Discriminatory uncertainty
  A model can discriminate not only through its decisions, but through how confident it is in them.
• July 2026 Why SHAP values are not points
  A SHAP waterfall looks like a scorecard. Read it like one and it hands out advice that is confidently wrong.
• August 2026 treecf: counterfactuals you can act on
  The only honest answer to “what would it take to be approved?” is a counterfactual — constrained, verified, fast.

Open source¶

probcal¶

python numpy-only MIT

Post‑hoc probability calibration for binary classifiers, built for regulated PD models. Three things scikit‑learn will not give you: logit‑scale diagnostics that keep a 1% default rate readable, a central‑tendency adjustment that ships as an auditable offset stage rather than a silent refit, and per‑grade regulatory backtests (binomial, Jeffreys). Around them: the calibration literature behind one API, exact inverse maps from policy cutoffs to raw scores, anytime‑valid monitoring, and JSON serialization. numpy is the only runtime dependency.

Documentation · PyPI · GitHub

treecf¶

python rust core MIT

Counterfactual explanations for tree ensembles (XGBoost, LightGBM, CatBoost, sklearn) that can prove their answer: the exact backend certifies that a plan is the cheapest available under the declared constraints, or that none exists. Targets are intervals on the raw output, so a plan moves a declined application under the cutoff the business actually uses. Plans respect declarative real‑world constraints, treat missing values as first‑class, and every answer is float‑verified against the parsed model before it is returned. The heuristic search runs on a Rust engine bundled in the wheel — milliseconds per applicant, whole portfolios in one parallel call.

Documentation · PyPI · GitHub

Earlier and experimental: concept-graph-xai (concept‑level feature usage and ablation), triadxai (three‑channel decomposition of credit scores), flaggam (interpretable GAMs built from univariate flags), swift (SHAP‑weighted drift monitoring).

Academic work¶

Doctoral thesis — Trustworthy credit risk under the AI Act¶

SGH Warsaw School of Economics in progress submission target 2029

Supervised by Prof. Aneta Ptak‑Chmielewska. The thesis develops an integrated methodology that combines algorithmic fairness, uncertainty quantification, and interpretability, organised around a concept I call discriminatory uncertainty — the idea that a model can discriminate not only through its decisions, but through how confident it is in them. The core idea is set out in the note Discriminatory uncertainty.

Conference talks¶

Slides and materials: github.com/wlazlod/conference-talks

Two of these talks are about knowledge graphs and ontologies. That is the same line of work seen from the other end: a shared vocabulary is a prerequisite for trustworthy ML, because you cannot audit the fairness or calibration of a feature nobody can define.

• The Scorecard Illusion: Why SHAP values are not points (and how they can fool you) — Data Science Summit, AI Edition 2026, Warsaw.
  Why a SHAP value behaves like a contribution, not a coordinate — and how reading it as a point misleads scorecard interpretation. Write-up.
• Curing Semantic Debt: Practical Applications of Knowledge Graphs in ML — GHOST Day 2026, Poznań, with Kamil Golis.
  How knowledge graphs pay down the semantic debt that accumulates across ML pipelines — a practical, applied walk‑through.
• Poza Feature Store: jak Ontologie Semantyczne mogą przyspieszyć rozwój modeli ML — Data Science Summit 2025, Warsaw.
  How semantic ontologies can move ML development past the feature store, giving models a shared, reusable vocabulary.

BibTeX entries for all talks: conference-talks README.

Teaching¶

Teaching materials: github.com/wlazlod/didactics

Data Mining — lecture course¶

SGH Warsaw School of Economics autumn 2026 taught in English

A finance‑first, problem‑driven Data Mining course for the SMMD‑ADA and SMMD‑AAB Master’s programmes (elective for three more). Seven blocks, a 102‑page coursebook, one lab notebook per class — and the whole course anchored on a single retail‑credit (PD) dataset with the teaching traps built in: three flavours of missingness and one deliberate leakage column.

Course script (HTML) · Repository

Conformal Prediction in Credit Risk — Research Workshop¶

Warsaw University of Technology with Allegro Pay

Industry‑mentored project for the Research Workshop course at the Faculty of Mathematics and Information Science. Three companies contribute a project each year and guide the student team through it; ours is on conformal prediction applied to consumer‑credit scoring. I co‑prepared the brief and mentor the team.

Starter repository · Teaching materials

Elsewhere¶