A showcase of my work. I enjoy identifying problems and finding great solutions. Scratching my own itch and hoping more people are feeling the same.
Rust-core IFC parser with a Python API — data out of building models in seconds, not minutes.
Opening an IFC and getting clean data out is painfully slow with the established kernels. ifcfast makes parse + query near-instant (~20–27× faster than ifcopenshell) by keeping geometry off the hot path.
Data-first BIM platform — files in, data streams out. Dashboards, verification and change detection are the product; the 3D viewer serves the insights.
People who use building models (not author them) have no tool where data, not geometry, is the unit of work.
Scriptable control of Solibri's geometry rule checking — rulesets, classifications, BCF, the full lifecycle over HTTP.
Solibri's geometry-aware checking is uniquely valuable and has no real automation surface. This makes it drivable by scripts and AI agents.
Self-service receiving-control of TFM tagging in IFC discipline models — upload, check, get an Excel + PDF report.
Verifying TFM labelling on incoming models is manual drudgery nobody has time for.
Survey files (CSV / Excel / KOF) to IFC4 site models in about thirty seconds.
Surveyed points normally wait on manual modeling before they exist in the coordination model.
This site is the demo: 9.6 million LiDAR points of Kyrkja, streamed in a purpose-built binary format (KYK3), materializing as it loads.
Photoreal isn't the only way to carry a place. Points — xyz and an elevation ramp — are lighter, stranger, and honest to the data.
A multi-machine AI fleet brain — three computers, one accumulating memory, coordinating over GitHub.
AI coding sessions forget everything between machines. The fleet should get tighter with every session.
Personal finance for Argentina's cash-and-three-currencies reality — statements, receipts, screenshots and voice notes in; multi-currency insight out.
No banking APIs, high inflation, ARS + USD + crypto side by side: the only viable aggregation is manual, so the product is the upload-to-insight loop.
Beat-reactive point-cloud visuals for DJ sets — Web Audio FFT driving starfields across multiple screens.
Live visuals without VJ software or canned loops — generated in the browser, in code.
Scenes in the register — places that matter, each with its own palette:
2,032 m in Jotunheimen. The church. Live now — you came in through it.
The 1994 Olympic downhill, drawn from the real World Cup course data.
The South Face, from the southern arc.
50 cm LiDAR — finer than the mountain that started this. Camera on the valley floor, looking up.
A forest tarn in the hills above Vinstra. Three kilometres, close to the ground.
Hjelmelandsdalen. A small ski hill, big untouched backcountry — and the farm the family name comes from.
How KYK3 works — delete-before-compress, Morton-shuffled streaming, render-on-demand, and an 8 GB no-GPU floor.
The edit → execute → assess → refine loop against a desktop app that never wanted one.