I’m Nathan — an ML Tech Lead and Architect based in Johannesburg, passionate about computer vision, clean code, and best practices.
This site will feature posts about machine learning, software architecture, open source projects, and other thoughts or ideas of mine.
Background
I’m currently at ByteFuse, working on AI and Agentic solutions for education, smart traffic systems, enterprise publishing tools, and R&D within embeddings and CV spaces.
Previously, I was Technical Lead at Dragonfruit AI, where I took ownership of their initial self-checkout fraud detection system, built scalable ML pipelines, oversaw development and improvement of many critical systems, and lead over 130+ technical interviews.
I hold an MSc in Computer Science from the University of the Witwatersrand, where my research focused on why disentangled representation learning fails within VAEs, often due to data and not the algorithms.
Research
Publications:
- Overlooked Implications of the Reconstruction Loss for VAE Disentanglement — IJCAI 2023
- MSc Thesis: Disentanglement Using VAEs Resembles Distance Learning and Requires Overlapping Data — Wits 2022
Open Source
norfair-rs — Multi-object tracking library in Rust. A high-performance port of the Python norfair library.
disent — Modular VAE disentanglement framework built with PyTorch Lightning, providing an efficient and easily comparable way to run and build VAE disentanglement models.
Experiments:
- pydependence — minimal pyproject.toml and requirements.txt dependency generation via static analysis over Python imports even across repos or from different entry points.
- mtg-vision — Real-time MTG card detection system for the real world using contrastive learning and completely synthetic data.
- mtg-dataset — Magic: The Gathering card image dataset with HDF5 support
- ruck — Evolutionary algorithms library with JIT and multiprocessing via Ray, providing massive speed ups for my MSc research.
- lightning-nca — Easy Neural Cellular Automata (NCA) in PyTorch Lightning.
Check back soon for updates.