Building practical machine learning and AI solutions for real-world problems. Specialising in predictive modelling, LLM engineering, and data-driven products with a focus on the South African context.
I'm Karabo, a data scientist and AI engineer based in Johannesburg — and these days my centre of gravity is machine learning and AI engineering in production. I designed, shipped and operated a full RAG question-answering system for Zaio: live web scraping with Playwright, LangChain chunking, Jina embeddings with reranking, Qdrant vector search, and a four-provider LLM cascade (Groq, Gemini, Together.ai, Mistral) answering real students' questions.
The classical ML record runs deep: credit scoring on South African banking data with SHAP explainability built for NCA/POPIA compliance, fraud detection handling severe class imbalance with SMOTE, Cape Town property price prediction, and a CNN for agricultural disease classification trained with transfer learning to 92% accuracy. I'm currently building EDITH — my personal AI agent on the Claude API with a full tool-use loop via MCP — and preparing for the ARC-AGI-2 reasoning challenge. Long-term target: Kaggle Grandmaster.
Alongside the research I build modern, responsive websites and web applications for clients who need a professional online presence — to the same standard I hold everywhere: clean code, honest documentation, and things that actually work.
See my workMachine learning, AI engineering, and full-stack projects — built to solve real problems.
Production RAG-powered Q&A API that answered Zaio students' questions from live-scraped site content. Playwright scraper, LangChain chunking (800/150), Jina v3 embeddings and reran…
Built a machine learning model to predict Airbnb listing prices in Cape Town using property features, location data, and amenities. Achieved 78% accuracy using Random Forest Regres…
Developed a credit risk assessment model for South African banking data. Used Logistic Regression and XGBoost to predict loan default probability, helping financial institutions ma…
Transaction fraud detection system built for South African banking patterns. Handles severe class imbalance with SMOTE resampling and XGBoost. High-recall model designed to flag fr…
Created a CNN-based image classification model to identify plant diseases from leaf images. Trained on 10,000+ images using TensorFlow and achieved 92% accuracy, helping South Afri…
A modern, responsive portfolio website built with PHP, MySQL, and vanilla JavaScript. Features a full admin panel for managing projects, skills, and content without touching code. …
Built across classical ML, LLM engineering, backend APIs, and full-stack development.
Available for data science consulting, ML engineering, AI product development, and web development projects. Based in Johannesburg, working with clients across South Africa.