Bedrock is a cloud-native LIDAR ground classification pipeline tuned for the failure case nobody else solves well: dense Southeast US canopy with sparse ground returns. Browser-first QC, COPC-streamed, surveyor-built.
Geometric filters (CSF, PMF, Axelsson TIN-densification) all share one failure mode: they assume "lowest local point = ground." In dense pine plantation, hardwood, cypress swamp, or kudzu, the lowest local point is often a root buttress or low understory. The DTM floats high and looks like a checkerboard of bumps. Surveyors spend 15–40 hours per heavy-canopy tract cleaning that up in LP360 or TerraScan.
v0 is the spike: prove the COPC-from-R2 → CesiumJS streaming architecture works, anchor the brand and UX, get a real lidar dataset rendering in a browser. The real pipeline lands in Phase 1.
CesiumJS streaming COPC over Cesium World Terrain + satellite. Try a public USGS 3DEP tile from Alabama.
Public COPC and LAZ sources we can stream today: USGS 3DEP, OpenTopography, DALES, Hessigheim, AWS Open Data.
PDAL → CSF → COPC → R2 → CesiumJS. Same boxes the v1 ML pipeline will fit inside.
The full scoping report — competitors, ML strategy, unit economics, phased plan, honest risks.
Cloud Optimized Point Cloud is a valid LAZ 1.4 file with octree indexing baked in. Range-read from Cloudflare R2 with $0 egress. CesiumJS handles the terrain context and mobile rendering. The viewer streams what's in frame, not the whole file. A 50 GB tract opens on an iPad.
$0.015/GB-month storage, $0 egress. A 500M-point LAZ at 1.5–3 GB streams free. AWS would eat your margin at $0.09/GB.
Won the streaming format war by 2024. USGS 3DEP, OpenTopography, IGN France, Sweden, NZ all standardized. Also a valid LAZ — every CAD reads it.
Current SOTA on every outdoor benchmark, with self-supervised pretraining that cuts labeling needs 3–10×. Fine-tunes from a checkpoint in 4–8 hrs on an A100.
QC edits stored as sparse Parquet delta, never rewriting the source COPC. Free undo, audit trail, PLS-signable history, 100× cheaper I/O than tile rewrites.
Every phase has a verifiable demo and a go/no-go gate. The internal-tool ROI is the floor. The SaaS upside is the option.
PDAL → CSF → COPC pipeline, public-data CesiumJS viewer, brand & scope landing site. Gate: pipeline runs end-to-end.
Job queue (D1), upload UI, DO GPU droplet running CSF + baseline ML, COPC+DTM+contours export. Gate: first paid Weygand job through.
Potree v2 embed, delta-edit layer (brush + polygon + undo), export with merged edits. Gate: in-house tech replaces LP360 on a real job.
PTv3 fine-tuned from Sonata, trained on 20+ Weygand tracts. A/B vs SMRF/CSF on 20 jobs. Gate: ≥95% canopy IoU AND ≥$30K/yr internal savings.
Multi-tenant auth, Stripe billing, share links, mobile viewer, Trimble Connect push. Gate: first paying customer outside Weygand.