Jean-Pierre van Zyl

About

Hi there, welcome to my personal portfolio.
I'm Jean-Pierre, a Stellenbosch University Computer Science PhD candidate. Working under the supervision of Professor A.P. Engelbrecht, my work explores the frontier of optimisation and machine learning, with a specific focus on making powerful technology both accessible and understandable. My current PhD thesis is focused on combinatorial optimisation advancements and the application to automated machine learning. Automated machine learning aims to democratise the power of modern day artificial intelligence, by making tools more readily available to non-technical users.

My Master's thesis, "Rule Induction using Swarm Intelligence", made a contribution in the field of explainable artificial intelligence with the idea that transparent models are more interpretable and therefore trustworthy. The thesis developed a new approach to extracting rules from tabular datasets by using the set-based particle swarm optimisation algorithm. The project contained multiple phases, including developing single-objective and multi-objective definitions of rule induction, as well as analysing the effect that different robust and non-robust classification metrics have on the approaches.

In my spare time I enjoy triathlon training, pretending that my problems don't exist, film photography, volunteering for SANParks, and making cocktails. The majority of my undergrad and masters degrees were profoundly involved with student leadership, through house leadership and student societies. If you want to know whether I want to create an AI that takes over the world, the answer is yes. But I'll settle for one which makes a lot of money.

Personal information.

Website: jeanpierrevanzyl.github.io
Phone: You'll have to buy me a drink first
Institution: Stellenbosch University
City: Stellenbosch, RSA

Occupation: Junior Lecturer, PhD candidate
Degree: MSc
Email: vanzylj -at- sun -dot- ac -dot- za
Freelance: Available

Resume

Hereby a summary of my education and previous work experience.

Summary

Jean-Pierre van Zyl

Brief overview of experience

Curriculum Vitae
Currently hold MSc in Computer Science
Work experience consists of multiple internships
Multiple years spent serving on student leadership

Education

PhD in Computer Science (In progress)

2023 -

Stellenbosch University, Stellenbosch, RSA

Currently studying under Professor A.P. Engelbrecht

Master of Science in Computer Science (Cum Laude)

2021 - 2022

Stellenbosch University, Stellenbosch, RSA

Obtained Cum Laude
View Degree

Bachelor of Science (Honours) in Computer Science (Cum Laude)

2020 - 2020

Stellenbosch University, Stellenbosch, RSA

Obtained Cum Laude
View Degree

Bachelor of Science in Mathematical Sciences (Computer Science)

2017 - 2019

Stellenbosch University, Stellenbosch, RSA

Completed a BSc in Mathematical Sciences with a main stream in Computer Science
View Degree

National Senior Certificate

2012 - 2016

Fairmont High School, Durbanville, RSA

Obtained 6 distinctions

Professional Experience

Junior Lecturer, Part-time

2025 -

Stellenbosch University, Stellenbosch, RSA (www.ie.sun.ac.za)

Lecturing and student supervision

Scholarship program

2023 - 2024

BMW IT Hub, Pretoria, RSA (www.bmwithub.co.za)

PhD funded by BMW as a contractor for the IT Hub

Teaching Assistant, Part-time

2022 - 2024

Stellenbosch University, Stellenbosch, RSA (www.ie.sun.ac.za)

Project marking and module admin

Intern, Part-time

Mar 2021 - Jun 2021

Merlynn Intelligence Technologies, Centurion, RSA (www.merlynn-ai.com)

Merlynn is an artificial intelligence firm developing customized AI solutions, without the need for historical data.
Worked remotely while studying my MSc, contributed to expanding the capabilities of their AI systems.

Intern, vac work

Jan 2021 - Feb 2021

Polymorph Systems, Stellenbosch, RSA (www.polymorph.co.za)

Polymorph is a software solutions company with wide range of focus and experience across industries
Worked as a part of the in-house R&D team which drove the expansion of the company's IoT work to include Computer Vision at the Edge

Intern, vac work

Nov 2020 - Dec 2020

NMRQL Research, Stellenbosch, RSA (nmrql.com)

NRMQL is an AI-driven investment company using cutting-edge technology to disrupt the industry
Position provided first introduction to the Fintech industry
Contributed to the expansion of the set of tools used to trade securities

Research Experience

All this stuff is fake. How does one even judge an open-ended skill as a percentage? But I see all the kids are doing it.

Data Science

Machine Learning

Particle Swarm Optimisation

Polynomail Approximation

Combinatorial Optimisation

Neural Networks

Rule Induction

Quantum Mechanics

Publications

Outlined are selected publications

Closed-Form Expressions for the Normalizing Constants of the Mallows Model and Weighted Mallows Model on Combinatorial Domains

This paper expands the Mallows model for use in combinatorial domains by deriving closed-form expressions for the normalizing constant of the distribution under various discrepancy functions, calculable in constant time. MDPI Mathematics.

Analysis of classification metric behaviour under class imbalance

This paper shows the unreliability of contemporary performance evaluation metrics for classification problems, and proposes a method to create robust metrics. Elsevier Egyptian Informatics Journal.

Set-Based Particle Swarm Optimisation: A Reivew

The main objective of this paper is to review the set-based particle swarm optimisation algorithm and to provide an overview of the problems to which the algorithm has been applied. MDPI Mathematics.

Rule Induction Using Set-Based Particle Swarm Optimisation

The SBPSO algorithm was successfully applied to induce rule sets on categorical data to allow for transparent classification for critical industries. This work has been published at the IEEE World Congress On Computational Intelligence (2022).

Polynomial Approximation Using Set-Based Particle Swarm Optimization

The SBPSO is an adaptation of the standard PSO for problems in a discrete search space. This algorithm was applied to solve regression problems and published for the International Conference on Swarm Intelligence (2021).

Student Supervision

A summary of past students supervised

Deep learning models, characterised by their multilayered complexity and vast parameter counts, have become the prevailing approach to image classification. Yet, the widespread adoption of deep architectures often occurs without critical assessment of whether such complexity is necessary. The study presented in this research assignment re-examines the assumption that deeper architectures are inherently superior. The investigation focuses on whether shallow neural networks, when supported by informed preprocessing, can achieve accuracy comparable to deep learning models while maintaining efficiency and interpretability. Three benchmark datasets of increasing complexity — MNIST, Fashion MNIST, and CIFAR-10 — are used to evaluate the hypothesis of the study. The study follows a structured, data-centric process which comprises exploratory data analysis, model-centric preprocessing, dataset-centric enhancement, and statistical validation. Class-specific preprocessing pipelines are developed to address residual misclassification, and incorporates handcrafted features such as edge, texture, and colour descriptors that reflect the visual attributes most relevant to each dataset. Across thirty independent runs, the optimised shallow networks achieve mean test accuraciesof 98.67% on MNIST, 88.16% on Fashion MNIST, and 78.44% on CIFAR-10. Performance on MNIST and Fashion MNIST match that of reported deep learning baselines, while the CIFAR-10 results demonstrate that shallow architectures remain viable for complex datasets when supported by targeted preprocessing, feature engineering, and ensembling. Statistical analyses confirm that improvements are significant and accompanied by reduction in required computational resources. The results challenge the presumption that depth is a prerequisite for high performance, and demonstrate instead that data-centric refinement can deliver competitive accuracy with substantially lower computational cost. The study contributes are producible framework for aligning model complexity with data characteristics, and advances the case for data-centric, efficiency-aware evaluation in modern image classification.

Andrea Mitchell

To Go Deep, Or Not To Go Deep

The research develops a particle swarm optimisation-based maximum-margin classifier that addresses the computational and theoretical limitations of quadratic programming used in training support vector machines. Conventional training in the primal or dual form requires strict convexity, satisfaction of the Karush-Kuhn-Tucker conditions, and clean, balanced data. In practice, real-world datasets often contain noise, class imbalance, and high dimensionality, which increase computational cost and reduce the reliability of quadratic program-ming solvers. To reduce computational effort and focus optimisation on informative samples, the procedure is restricted to Tomek link pairs located near class boundaries. The removal of redundant and non-support vectors lowers the effective problem dimensionality and improves margin estimation, which yields more stable and generalisable decision boundaries. The framework is extended to non-linear classification through kernel functions, such as radial basis function and polynomial kernels, which implicitly map data into higher-dimensional feature spaces to model complex decision surfaces under class overlap and noise. Experiments on seven benchmark datasets and four synthetic datasets show that the proposed classifier achieves classification accuracy and F1 scores comparable to, or exceeding, those of support vector ma-chines trained using quadratic programming. Although the method requires greater computational resources due to iterative fitness evaluations, it provides a flexible, solver-free alternative that avoids assumptions of convexity, differentiability, and dual formulation.