Tiaan van Antwerpen
I am an AI engineer at Spatialedge, where I build production-grade tools that fit into how people already work. I usually work forward-deployed, sitting with the client from the first prototype through to production.
About
I come from an industrial engineering background, which taught me early to think in systems: how people, processes and technology intersect to create value, or quietly break it. That mindset stayed with me as I moved into machine learning, and it is why I tend to treat the model as one component among many. It is rarely the component that decides whether a project succeeds. The handoffs between steps, the quality of the underlying data, and whether anyone trusts the answer enough to act on it usually matter more.
My master’s at Stellenbosch was on adaptive LLM routing for agentic workflows, completed as a case study with Spatialedge and released as an open-source framework. The idea is to select an appropriate model per query so that a workflow keeps its accuracy without paying for a frontier model on every call.
I am currently a consulting AI engineer at Spatialedge, working forward-deployed on end-to-end enterprise solutions. Recent projects include a diagnostic system for a fleet of more than forty thousand vehicles, a claims processing platform for a transport company in the United States, and enterprise search. The work is not about chasing whatever model is trending this week; it is about building production-grade tools that fit into a real workflow and make someone’s workday noticeably easier.
Work
Smaller experiments are on GitHub.
Writing
Notes on routing, evaluation and enterprise AI.
Mostly the space between a model that scores well on a benchmark and a system someone is willing to rely on.
The model is not the hard part
Notes from forward-deployed enterprise AI work: the model is one component among many, and rarely the one that decides whether a project succeeds.
Read itAdaptive LLM routing for effective agentic workflows
A summary of my MEng research: routing between thirteen models in a text-to-SQL agent, what worked, and why none of it beats simply picking a good model.
Read itBackground
AI & Software Engineer, Spatialedge
I built the first proof of concept for DataLlume, the company's enterprise GenAI product, and owned its early development as it grew into a production system now used at several large corporates. Since then I have delivered a claims processing platform for a transport company in the United States and an agentic diagnostic system for a fleet of more than forty thousand vehicles, both forward-deployed with the client.
MEng Data Science, Stellenbosch
Completed cum laude. My thesis, Adaptive LLM Routing for Effective Agentic Workflows, was done as a case study with Spatialedge.
Industrial Engineer in training, Blackslope
I evaluated early-stage startup ventures through market analysis, value chain assessment and competitor research, and helped shape their initial value propositions.
BEng Industrial Engineering, Stellenbosch
A four-year generalist engineering degree covering operations research, mathematical modelling and systems design. My research project built a decision support system that used computer vision to classify pesticide decomposition on leaves.