Construction geometry.

Donovan Group builds two quantity surveying platforms on the same model: uTecture for residential plans, and AirBuildr for commercial metal buildings — warehouses and other portalised steel structures. A design goes in, the platform builds a 3D model from it, and that model is measured against a materials database to produce the take-off and a price that updates as the design changes. The 3D view is a preview of the building being priced.

That puts the geometry directly underneath the numbers. Areas and volumes measured off the model are what the estimate is computed from, so a triangulation that degenerates on an awkward footprint is not a visual artefact in the preview. It is the wrong quantity of framing or cladding in a quote somebody is about to bid on.

Adeline's work sat evenly across both platforms' backend, a large Java and Spring codebase: polygon triangulation, 3D shape handling and automatic floor-plan/model generation, plus algorithm and data-processing optimisation across the generation pipelines.

She also kept the automated tests and handled production bugs. In a geometry pipeline that mostly means tracking down the exact input that broke, usually one awkward footprint from one job, and reproducing it locally.

The same triangulator and the same input problems sat behind both uTecture's residential plans and AirBuildr's commercial designs. The biggest source of difficulty was the input, not the algorithm: source building models were best-effort rather than volumetrically sound, with measurements missing or not making sense on their own terms. That, and doing intensive geometry work in Java, where memory was a constant constraint. The codebase itself was legacy: no single coding convention, and years of contributions written in whatever style each developer favoured at the time.

A correct floor plan was an airtight volume with regular triangulation — no degenerate triangles, and not subdivided further than it needed to be — validated both visually and mathematically. The take-off itself measured a mix of derived quantities, because a clean source of truth often didn't exist; an exact floor outline, for example, usually had to be derived through a chain of rules rather than read directly off the input.

The original triangulator was fast because it ignored constraints, and it showed: visually broken output on anything but a simple footprint. The constraint-respecting rewrite was correct but roughly three to four times slower.

Geometry tests asserted polygon count, edge matching, point connectivity and usage, dimensions and area, backed by native shape and area class extensions for converting between different shape representations quickly and accurately.

Beyond the triangulator, she did general front-end, API and backend work on the platform, and late in the role redesigned how the platform generated manifold roofs — the existing approach wasn't producing manifold geometry despite the name.