3D reconstruction expert - hole-filling / mesh completion for outdoor phone scans

What we do We reconstruct metric-accurate 3D of outdoor scenes (yards, driveways, rooftops) from ordinary iPhone video - RGB, sometimes with LiDAR. The output feeds real measurements, so geometry has to be true, not just good-looking. The problem Our reconstructed point clouds and meshes have holes — coverage gaps, depth dropout on thin/dark/distant surfaces, open TSDF/Poisson boundaries, and ground removed by over-eager segmentation. We want these closed without fabricating geometry. A fill that invents surface area is worse than the hole, because it corrupts downstream measurements. Closing real gaps confidently while clearly leaving genuinely-unknown regions unfilled is the core skill we're testing for. Stack (yours to work within) Python, Open3D, COLMAP, gsplat, TSDF + screened Poisson meshing, monocular metric depth (Depth-Anything family), SAM2 segmentation, GLB/Draco output, Modal for serverless GPU. You don't need every one of these, but you must be fluent in point clouds, depth maps, and meshing. What a great outcome looks like Measurably fewer / smaller holes on our test scans No inflation of measured surface area vs. a known reference A principled distinction between "confidently filled" and "unknown/left open" (e.g. per-vertex confidence) Approaches we can run in a batch pipeline, not a one-off manual clean-up To apply, please: (1) link one reconstruction you personally built end-to-end where you dealt with holes/incomplete geometry, (2) name the specific technique you'd reach for first here and why, (3) how you'd measure whether a fill is honest vs. fabricated.