GA Gaël Aglin
Senior Data Scientist · Ph.D.

Gaël
Aglin.

Production ML · Decision Science · GenAI

I build production decision systems that combine machine learning, causal experimentation, and combinatorial optimization, backed by measured impact and widely cited research.

Get in touch Google Scholar GitHub
↳ Brussels, Belgium ↳ Open to relocate ↳ 8+ years
An optimal decision tree. The accent path is the proven-optimal route from root to leaf, the core idea behind DL8.5 (AAAI 2020).
01

Selected record

Public, verifiable figures. Research and open-source metrics, competition results.

218

Citations, AAAI 2020

“Learning Optimal Decision Trees Using Caching Branch-and-Bound Search.” Presented at AAAI, New York.

290 / h-index 4

Total citations

Across AAAI, IJCAI, ECML PKDD, IDA, and CPAIOR. Tracked on Google Scholar.

21 releases

PyDL8.5, open source

scikit-learn-compatible C++/Cython/Python library on PyPI. 600+ commits, CI on Linux, macOS and Windows.

1st

BNP Paribas Group RAG Hackathon 2024

First place in a multi-team Retrieval-Augmented Generation competition across 12 country offices.

1st

Proximus Kaggle Competition 2023

First place. Object detection / image classification on industrial fiber imagery.

02

Profile

Eight years turning hard decision problems into systems that ship, where machine learning meets causal inference and combinatorial optimization.

I'm a Senior Applied Data Scientist with a Ph.D., working across predictive ML, causal experimentation, and constrained optimization. At BNP Paribas Fortis I own end-to-end production decision systems for the pricing business line, from price-elasticity and risk-cost modeling to constrained optimization, validated through A/B testing before rollout.

My research on optimal decision trees (DL8.5) was published at AAAI 2020 and is widely cited, and I maintain the open-source PyDL8.5 library. I've led industrial collaborations on steel-sheet production scheduling (PSI Metals) and inventory forecasting (H&M Sweden), and I serve on my bank's AI Model Validation Committee.

03

Experience

Industry and applied research, in parallel for most of the last decade.

Industry

2023 – now

Senior Data Scientist BNP Paribas Fortis · Brussels, Belgium

  • Own end-to-end production decision systems spanning predictive ML, causal experimentation, and constrained optimization for the pricing business line (retail / private / corporate banking); drive cross-functional execution with risk, finance, IT, product and Responsible AI under GDPR and the EU AI Act.
  • Designed and shipped a production dynamic-pricing framework combining price-elasticity / uplift modeling, probability-of-default and risk-cost prediction, and constrained non-linear optimization, validated through A/B testing, with significant measured business impact.
  • Run production model monitoring (performance, drift, business KPIs); when drift appears between A/B groups, isolate the true causal effect via post-randomization bias correction and causal inference (matching, difference-in-differences).
  • Architected and productionized an intelligent document-processing system using Retrieval-Augmented Generation for credit-acquisition workflows, with evaluation strategies and guardrails for LLM outputs.
  • Member of the bank's AI Model Validation Committee: model-governance contributor and technical reviewer, from use-case framing through pre-production sign-off.
Causal inference Uplift modeling Constrained optimization Production RAG A/B testing
2018 – 2020

Data Scientist Consultant bpost · Belgium · part-time, alongside Ph.D.

  • Built NLP-based document classification for automated postal document processing (structured PDFs, scanned forms, document images).
  • Delivered geospatial and spatio-temporal analytics on postal logistics (flows, routes, volumes) with GeoPandas to support operational decisions.
2016 – 2017

Android Developer HE Systems · Cotonou, Benin

  • Built and shipped an e-commerce Android application end-to-end, early production-system experience.

Research

2022 – 2023

Postdoctoral Researcher, ML & Combinatorial Optimization UCLouvain · Belgium

  • Led a one-year industrial collaboration with PSI Metals (Brussels): replaced a genetic-algorithm baseline in their steel-sheet production-scheduling engine with a Structured Perceptron preserving sequence-structure constraints; published at CPAIOR 2024.
  • Developed ML-enhanced optimization frameworks for decision-making under preferences; supervised Master's theses and mentored students.
2018 – 2022

Ph.D. Researcher, ML & Combinatorial Optimization UCLouvain · Louvain-la-Neuve, Belgium

  • Advisors: Prof. Siegfried Nijssen, Prof. Pierre Schaus. Designed exact algorithms for optimal decision trees (DL8.5), outperforming prior MIP approaches by orders of magnitude (AAAI 2020).
  • Led an industrial collaboration with H&M Sweden: adapted optimal decision trees into quantile regression trees for inventory min/max forecasting on Databricks.
  • Released and maintained PyDL8.5 (scikit-learn-compatible, PyPI). Teaching assistant in Python programming, databases, AI, and algorithm complexity.
04

Research & publications

Selected peer-reviewed work. All publications →

AAAI2020

Learning Optimal Decision Trees Using Caching Branch-and-Bound Search

Aglin G., Nijssen S., Schaus P.

Presented at AAAI, New York.

218 cites
IJCAI2021

PyDL8.5: A Library for Learning Optimal Decision Trees

Aglin G., Nijssen S., Schaus P.

49 cites
IDA2024

Interpretable Quantile Regression by Optimal Decision Trees

Lemaire V., Aglin G., Nijssen S.

CPAIOR2024

An Efficient Structured Perceptron for NP-Hard Combinatorial Optimization Problems

Véjar B., Aglin G., et al.

Industrial collaboration with PSI Metals on steel-sheet production scheduling.

05

Selected projects

Open source · 2019 – now

PyDL8.5

Library for inferring optimal decision trees, scikit-learn-compatible. C++ core (~74%), Cython bindings, Python wrapper. Published on PyPI, documented on ReadTheDocs, CI across Linux, macOS and Windows.

C++Cythonscikit-learnPyPI
github.com/aia-uclouvain/pydl8.5
Solo B2B SaaS · 2024 – now

rendoc.africa

End-to-end B2B SaaS platform (pre-production). Next.js full-stack, PostgreSQL + Redis + S3-class storage, deployed on Railway behind nginx on a Hetzner VPS. Embedded ML: face detection + matching for ID verification (KYC), topic modeling and sentiment on reviews. Brevo email and WhatsApp Business via Twilio.

Next.jsPostgreSQLRedisML / KYC
rendoc.africa
06

Capabilities

ML, Causal & Optimization

A/B testingUplift modelingDifference-in-differencesMatchingBias correctionDrift detectionQuantile forecastingPrice-elasticityConstraint programmingBranch-and-boundMIP · Gurobi · CPLEXStructured Perceptron

GenAI / LLMs

Production RAGIntelligent document processingLangChainPrompt engineeringAgent / tool orchestrationEval & guardrails

Programming & Cloud

PythonSQLC++RJavaPyTorchTensorFlowscikit-learnPandas · NumPyCythonDatabricksPostgreSQL · RedisAWS (S3, SageMaker, Bedrock)

Engineering practice

CI/CD · GitHub Actionspre-commit · ruffCode reviewModel cardsProduction monitoringnginx
07

Education

2018 – 2022

Ph.D., Engineering Sciences (Computer Science) UCLouvain · Louvain-la-Neuve, Belgium

  • Artificial Intelligence & Combinatorial Optimization. Thesis: Optimal Decision Trees Under Constraints.
2014 – 2016

M.Sc., Information Systems & Computer Networks IFRI · Cotonou, Benin

  • Graduated with Distinction.