FlowState — CYVL Hackathon

Overview

CYVL's constraint: build something useful with their mobile-LiDAR scans of city streets. FlowState is the flood-risk map that resulted, showing the public where water will go during a storm. I built the pipeline in between, turning CYVL's raw 354 MB street scan (hundreds of millions of points) into both a 3D model you can view in a browser and the terrain data the flood simulation needs. That included writing a ground-classification step from scratch after the standard tool wouldn't install mid-hackathon.

The Pipeline

One input, two outputs:

1. Raw mobile-LiDAR .laz scan of city streets — 354 MB, hundreds of millions of points.
2. SMRF ground classification separates ground returns (road, sidewalk, terrain) from everything above it (cars, poles, canopy).
3. Fork one: ground points become contour lines, exported to DXF for AutoCAD / Civil 3D — the format engineers actually work in.
4. Fork two: ground points become a heightfield mesh, exported to OBJ/MTL, uploaded to Autodesk Platform Services (signed S3 upload → SVF2 cloud translation), and rendered in a browser-based Viewer.
5. That same classified geometry is the input to the flow simulation whose output FlowState displays.

Tech Stack

• Point cloud processing: PDAL (SMRF ground classification), with a hand-written NumPy fallback
• Geometry export: contour lines → DXF, ground mesh → OBJ/MTL
• 3D viewing: Autodesk Platform Services — signed S3 upload, SVF2 translation, browser Viewer SDK
• Frontend: Next.js map UI (FlowState) consuming the simulation output

My Contributions

Solo repo — every commit in the pipeline is mine, built in one ~8-hour day.

• Wrote the full point-cloud pipeline: header-only extent reads, PDAL-side cropping, 0.10 m voxel downsampling, and vectorized NumPy scatter-adds (np.add.at) in place of Python loops to keep hundreds of millions of points tractable on a laptop
• Built the contour → DXF export path for AutoCAD / Civil 3D
• Built the ground mesh → OBJ/MTL → Autodesk Platform Services pipeline: signed S3 upload, SVF2 translation, and the browser Viewer integration
• Fixed a WebGL precision bug — raw UTM 19N coordinates (~328000, ~4692000 m) break float precision on the GPU, so geometry is origin-shifted to the region-of-interest centre before export

Challenges

PDAL is conda-only and refused to install on an M3 Mac mid-hackathon. Rather than lose the classification step, I wrote a from-scratch NumPy replacement for SMRF ground classification: sort every point by (elevation, grid-cell key) with np.lexsort, find each cell's first occurrence with np.unique(return_index=True), then pull a per-cell 10th-percentile elevation with pure offset arithmetic — no Python-level loop over hundreds of millions of points.

What Didn't Ship

Road / sidewalk / grass material classification is fully built — a 14-material MTL, per-cell argmax voting, and primitives for cars, poles, hydrants, and canopies — but the classifier never got wired into the final run: run.py hardcodes every ground cell to "grass". The Somerville ArcGIS fetcher that would have supplied real road and sidewalk polygons was written and never called. Both are scaffolding for a v2 the clock didn't allow.

Project Media

Links

Tools & Methods

Built at the CYVL Hackathon.