The hard problem
Raw depth scans from a phone are noisy, holey and fragmented — and a custom orthotic is manufactured directly from the scan. The output has to be one clean, watertight, dimensionally accurate mesh, produced in seconds, by a non-technical operator, on consumer hardware.
That last mile — from raw point soup to a manufacturing-ready model — is where most of the difficulty lives, and it's the part I owned.
What I built
- The complete white-label app — built the full application around the scanning framework (scanning flow, UI, uploads, brand configuration) and shipped it to the App Store as 12+ branded apps — DigiFoot, Pedico, Flippi, ped3D, Vesta, Feeting, FOUZIES, HKS SOLEMATE, Vôtres, WearSun, Avidor — and still growing, with new brands onboarding and thousands of scans processed across them.
- Cut scan-to-mesh from 3–5 minutes to under a minute. The mesh-cleanup pipeline was a JavaScript path taking 3–5 minutes per scan; I rewrote it in C++/Metal and embedded it directly inside the scanning SDK, so a finished, watertight mesh comes out in under a minute — seamlessly, with no extra step for the operator.
- Doubled measurement accuracy — tightened foot-measurement accuracy from ±10 mm to ±5 mm or better, the difference between a usable scan and a manufacturable orthotic, at a 30 fps live preview.
- Reconstruction pipeline work — the depth-frame → 3D model build on top of ICP alignment and surfel fusion (C++ core, Metal-accelerated preview).
- Surface smoothing — tuned to remove sensor noise without eroding the anatomical detail the insole geometry depends on.
- Meshing & export — point cloud → watertight mesh (Poisson reconstruction) → STL plus measurement metadata consumed by insole CAM tooling.
The scanning stack builds on the open-source StandardCyborgFusion engine; landmark detection was built with the wider team.
How the system works
- TrueDepthRGB + depth @ 30 fps, live quality checks
- FusionICP alignment + surfel fusion (C++/Metal)
- Cleanupoutlier removal, consolidation, smoothing
- MeshPoisson reconstruction → watertight STL
- Manufacturemeasurements + metadata → insole CAM
Scanning runs two paths in parallel: a GPU-rendered live preview so the operator sees the model forming in real time, and a CPU reconstruction that accumulates the final 10–25k-point cloud. Quality gates (distance, tilt, lighting, alignment) guide the operator before a bad scan ever reaches the mesh stage.