Applied AI · Data Science · AI Engineering · Research · Founder of Ubunye AI Ecosystems South Africa · from Soshanguve

I build machine learning and data systems that run in production, and I write about how. Nine years across insurance, telecommunications, applied research and higher education: enterprise ML platforms, Spark and Databricks pipelines at scale, and the unglamorous engineering that keeps models alive after the notebook is closed. That experience pushed me toward reusable infrastructure and applied research.

I am preparing doctoral research at the University of the Witwatersrand on physics informed self supervised learning for SAR based flood extent mapping: remote sensing, self supervised learning and computational hydrology, aimed at insurance risk in data scarce regions. (Proposal stage; not yet registered.)

Everything I build in the open goes through Ubunye AI Ecosystems, an organisation for producing serious open source from Africa rather than only consuming it.

Agent Roadmap Kit: the free, hands on companion to my six part series on building with AI agents. Six stops, one working assistant that answers questions from your own notes, with tools, search by meaning, safety, tests and CI. Runs on Gemini's free tier, a local model through Ollama, or fully offline. Read the series.

Ubunye Engine: a config driven, Spark native framework for data and ML pipelines. Describe a pipeline once as a folder of config plus Python, then run that exact folder on a laptop, Docker, Kubernetes or Databricks. A build job runs one pipeline on five execution environments and fails if the outputs differ. pip install ubunye-engine

tfilterspy: Bayesian filtering in Python. Kalman, Particle and Ensemble filters behind a scikit learn style API, scaling out with Dask. For anyone estimating true state from sensors that lie.

Ubunye Examples: worked, deployable pipelines on Databricks. The proof the framework does what it claims.

Ubuntu CFD: research. Cost function discovery under Ubuntu constraints: an agent that discovers cost functions rather than optimising a fixed objective, arguing that fixed extractive objectives are mathematically destabilising over long horizons, not only ethically wrong.

Echo State Networks for image segmentation: code behind my MSc dissertation at Wits, learning variational level set segmentation as a spatiotemporal problem. The research, line by line.

Published work on hearing loss estimation in mine workers, noise policy systems for mining, and long range seasonal temperature forecasting. Full list, with BibTeX for every paper: tmashininisekgoto.com/publications

Kasilam Digital Platforms: a free community initiative where young people in South African townships learn to build with AI by shipping real sites for local businesses. I teach; the participants build.

I write about AI engineering, MLOps, Spark and Databricks, and building systems in South Africa, including The Practical Roadmap to Building With AI Agents, a six part series where every post ends with something you can run.

tmashininisekgoto.com/blog · also on dev.to · RSS

_{Working with: Python · Apache Spark · Databricks · PyTorch · Kubernetes · MLflow · Airflow · TypeScript}