Work

General-purpose language models are mediocre readers of economic text: the sentiment of a monetary policy statement is not the sentiment of a movie review. This work finetunes LLMs on domain text so they pick up what actually matters — hawkish versus dovish tone, stance shifts, the difference between describing inflation and worrying about it.

It is the production-scale continuation of my PhD research, which built transformer-based methods (including a custom similarity-transformer architecture, Simtran) for classifying small, dense corpora of central bank communication.

Research data centres give researchers controlled, secure access to confidential microdata — the kind of granular administrative data that transforms what empirical economics can answer. I work on building this capability at the South African Reserve Bank: the infrastructure, the data pipelines, and the engineering that makes sensitive data usable for research without compromising it.

I’ve worked on the other side of this fence too — as a research assistant at UNU-WIDER’s SA-TIED programme I worked with the SARS administrative tax data, so I know exactly what researchers need from an RDC.

Beyond single-shot prompting: orchestrating LLM agents into workflows that hold up in production — multi-step pipelines for document processing, research support, and data engineering. The interesting problems are the unglamorous ones: evaluation, guardrails, structured outputs, and knowing when a deterministic script beats an agent.

Economic data increasingly arrives as pixels: scanned documents, forms, satellite imagery, photographs. I finetune computer vision models to turn that into structured, analysable data — the visual counterpart to the NLP work, and part of the same thesis: most of the information economists need is locked in unstructured formats.

My PhD work at the University of Pretoria, under the SARB Chair in Monetary Economics: applying computational linguistics to South African Reserve Bank communication. Sentiment, time, topic and consistency modelling of MPC statements; linking policy language to market reactions; and Simtran, a similarity-transformer architecture for classifying small, dense economic text corpora.

The work lives in the UP Macrolab’s central bank communication workstream, where I remain a Research Fellow.