What we were trying to build
The idea sounds simple: point an iPad at a construction site, see the BIM model overlaid live through the camera, and catch discrepancies before they become expensive. Anyone who has actually tried to build this knows the phrase "sounds simple" is doing enormous work.
A construction site is one of the hardest environments in the world for AR. The lighting changes hour to hour. The dust and debris confuse depth sensors. The GPS signal drops under a slab. The model has hundreds of thousands of elements. And the user needs it to work while walking, one-handed, in a hard hat.
DeltaARBIM is our answer to that problem. Here's what we had to solve to get it working.
Challenge 1 — Aligning a huge BIM model to a physical site with ±2 cm accuracy
Standard AR frameworks like ARKit can track a phone's position in a room accurately, but a construction site isn't a room. It's a moving target with reflective surfaces, unmarked landmarks, and a floor plate that changes weekly.
We built a hybrid anchoring pipeline that combines three sources of truth:
- iOS LiDAR scanning to capture a dense point cloud of the visible geometry
- GPS + visual anchors to lock the model's origin to real-world coordinates
- Continuous SLAM correction using ARKit's world tracking to keep the overlay stable as the operator moves
The three signals cross-check each other. When one drifts (LiDAR loses contrast, GPS weakens indoors, SLAM accumulates error), the other two pull the alignment back. In field tests, the resulting overlay stays within ±2 cm of the model position — good enough for MEP clash verification, structural check-off, and installation validation.
Challenge 2 — Rendering a Revit model with millions of polygons at 60 FPS on a mobile device
A real BIM model isn't a coffee-table 3D scene. A single hospital floor plate in Revit can hit 3 to 5 million polygons across all disciplines. Naive rendering would kill an iPad in ten seconds.
Our rendering pipeline does three things aggressively:
- Pre-processes the model through a custom converter that decimates non-essential geometry, merges coplanar surfaces, and groups elements by discipline
- Streams only what's visible using frustum culling, occlusion culling, and distance-based level-of-detail
- Renders layer-by-layer so the operator can toggle structure, HVAC, water, electrical, or fire independently — only the active layer is drawn full-detail
The result: sustained 60 FPS on iPad Pro and iPhone Pro, even on floor plates that would choke a desktop workstation running Revit natively.
Challenge 3 — Supporting IFC, Revit, and Navisworks without breaking the BIM workflow
The AEC industry doesn't work in one format. Structural engineers hand over IFC. MEP consultants ship Revit. Coordinators produce Navisworks federated files. A tool that only accepts one of these is dead on arrival.
We built our importer around IFC as the canonical format, with dedicated Revit (.rvt) and Navisworks (.nwd) converters on top. Every model — regardless of source — passes through the same optimisation and metadata-preservation pipeline. Element GUIDs, parameters, and classification data survive the round-trip, so an issue flagged on site links back to the exact element in the source model. No manual re-tagging.
Challenge 4 — Making it usable one-handed, in a hard hat, in bright sun
Engineering isn't only backend work. The last challenge — and often the hardest — was designing an interface a site engineer could actually use while walking a live construction floor.
We stripped the UI to four essentials: layer toggles, an alignment reset, a discrepancy flag, and a photo capture. Everything else lives one tap deep. Buttons are oversized. Contrast is tuned for outdoor daylight. The alignment reset is a two-finger gesture so it can't be triggered by accident. Every interaction has haptic feedback because the user probably can't hear the tap over an angle grinder.
Field-tested with real site engineers before the first line of the production build was written.
What it all adds up to
Individually, none of these problems is unique. Together, they define the difference between "a demo that works in an office" and "a tool that survives on a live construction site." That's the line DeltaARBIM was built to cross — and the reason our team spent three months in R&D before shipping the first stable release.
The next challenges are already on the roadmap: multi-user shared AR sessions, real-time model sync with cloud CDEs, and predictive clash analysis using on-device machine learning. Each one is a technical case study of its own.