We're building a drone LiDAR intelligence platform that continuously scans residential rooftops across Denver metro and scores them for hail damage and age using AI. We fly DJI Matrice 400 drones with Zenmuse L3 LiDAR sensors at 150m altitude. DJI Terra Pro processes the raw L3 data into classified .LAS point clouds. Your job starts there. This is a focused 6–8 week contract with a defined deliverable. We have a CTO who will integrate your output into our production stack — you're building the model, not the surrounding infrastructure. We need someone who has worked with LiDAR point cloud data before. This is not a learning opportunity. What you're building: 1. Roof segmentation pipeline Takes classified .LAS output from DJI Terra (Zenmuse L3, 150m AGL, Denver residential) and isolates individual roof polygons per parcel. Must handle: Complex rooflines, dormers, hip roofs, gable roofs Attached and semi-detached structures Tree occlusion and partial canopy interference Dense urban and suburban residential environments Output: per-property roof polygon with plane-level segmentation, tied to parcel boundary via geospatial join. 2. Hail damage and age classifier Scores each roof plane on: Hail damage severity (none / minor / moderate / severe) Estimated remaining useful life Material type where determinable from point density and return characteristics Confidence score per assessment Output: property-level damage score with plane-level breakdowns and confidence weighting. 3. Validation framework Accuracy metrics and confidence thresholds validated against ground truth. We have access to: Historical contractor inspection reports and insurance claim files from 25+ Denver roofing partners HailTrace property-level hail history for Denver metro Before/after point cloud diffs from storm response flights (available Month 3+) Your model needs to be accurate enough to be defensible to insurance carriers by Year 2. We're not asking for carrier-grade actuarial precision on Day 1 — we're asking for a solid v1 that improves as our proprietary dataset grows. 4. Retraining pipeline The model will be retrained continuously as our proprietary data accumulates — contractor inspection reports correlated with scan data, post-storm before/after diffs, carrier claims data. Build it to be retrained, not frozen. 5. Clean handoff to CTO Documented code, retraining guide, 2-hour technical handoff session. Our CTO is a strong AI systems engineer — not a geospatial specialist. He needs to own this after you're done. Training data you'll have access to: Available immediately: USGS 3DEP LiDAR coverage for Denver metro — free, download today ISPRS benchmark datasets for building detection and roof segmentation Published academic labeled roof datasets (IEEE/MDPI roof classification dataset — 4,500+ labeled roofs) HailTrace property-level hail event history for Denver — storm date, size, location Historical contractor inspection reports and Xactimate claim files from Denver roofing partners Available Month 3–4: Our actual Zenmuse L3 scan data from Denver baseline flights Before/after point cloud diffs from first storm response flights Contractor ground-truth validation of scored properties Your first 4–6 weeks use public and partner data. Final tuning uses our proprietary L3 data. The model you deliver will be retrained on increasingly rich proprietary data after handoff — that's by design. Required experience: Point cloud processing — PDAL, Open3D, or equivalent Building and roof segmentation from airborne LiDAR ML model training and validation in Python Geospatial data formats — .LAS, .LAZ, GeoJSON, GeoTIFF, WKT At least one segmentation framework — PointNet++, RandLA-Net, or similar Geospatial joins and parcel boundary matching Nice to have: Experience with drone-acquired LiDAR specifically vs terrestrial or manned aircraft Familiarity with hail damage signatures in LiDAR return characteristics Experience correlating point cloud data with ground-truth inspection reports Insurance, proptech, or roofing industry background PDAL pipeline automation experience What we're NOT looking for: RGB photogrammetry specialists without LiDAR experience General ML engineers without geospatial background Anyone who needs time to learn point cloud processing during the contract Anyone proposing to use Nearmap or CAPE Analytics — we are building LiDAR-native, not RGB-based scoring Deliverables: Roof segmentation pipeline — Python, documented, production-ready Damage and age classifier v1 — trained, validated, documented Validation report — accuracy metrics, confidence thresholds, known limitations Retraining pipeline and guide for CTO 2-hour technical handoff session with CTO Milestone payment structure: Milestone 1 (Week 2): Roof segmentation pipeline working on USGS 3DEP Denver data — 25% Milestone 2 (Week 5): Damage classifier v1 trained and validated on available datasets — 50% Milestone 3 (Week 7–8): Final tuning on L3 data, handoff complete — 25% Engagement: 6–8 weeks 35–40 hrs/week 100% remote Start: immediate Milestone-based payments To apply, include: One specific LiDAR project you've worked on — input data, what you built, accuracy achieved Your preferred tech stack for roof segmentation from airborne LiDAR Your approach to training a damage classifier when labeled hail damage LiDAR data is limited Your hourly rate Applications without a specific LiDAR project example will not be reviewed.