I take AI products from whiteboard to production — and keep them there.
For the past four years I was a Staff Machine Learning Engineer at Wunderkind, where my team's models decided the send time of roughly 430,000 marketing messages a day, on infrastructure processing over 2 trillion events a year. I built the Transformer-based timing model that lifted conversion rates 17%, the simulator we used to evaluate models offline before risking live traffic, and the experimentation framework that carried nine model generations into production. Along the way I led a team of three ML engineers across four initiatives.
Before that: machine learning for drug-target discovery at BioAge Labs, where the platform I built supported the in-licensing of three clinical-stage molecules, and earlier, backend and blockchain systems for European public-sector clients.
Now I'm building ConstructOS — configure-to-order software for building-product manufacturers (steel halls, timber frames). One data model carries a project from design brief to production files: parametric 3D configuration, engineering rule validation, pricing, quotation, BOM. It exports clean IFC 4 (zero validation issues in IfcOpenShell), plus DXF, STEP, and glTF, straight from the browser. I designed and shipped it solo: nine services, ~2,700 automated tests, live today at
constructos-9ul.pages.dev.
I'm currently talking to building-product manufacturers interested in piloting, and to investors who back technical founders in construction tech.
If that's you — or if you want to compare notes on ML systems, BIM interoperability, or the jump from staff engineer to founder — my inbox is open.
