RoadScan

• 

  clustered map view

• 

  survey dashboard

• 

  initial form to submit

• 

  capture screen

• 

  manual attachment dialog box

• 

  report

• 

  result and analysis of a capture

Inspiration

Canada has a $357 billion municipal infrastructure deficit. Federal and provincial grants exist to fix it, but they require a formal asset condition assessment before a town can even apply. That assessment costs \$30–50k. Small towns skip it, never apply, and the money sits unspent while their roads fall apart.

The FCM puts it simply:

$$ C_{\text{deferred}} = 7 \times C_{\text{preventive}} $$

Every dollar you skip today costs seven within five years. And every small town has a road foreman who knows exactly where the bad roads are — but that knowledge lives in his head and disappears when he retires.

RoadScan closes both gaps. Drive the roads, capture damage, and the output is the grant application. Not a cheaper survey, a grant eligibility engine.

Challenges I Faced

• Local YOLO was unusable. YOLOv8 on a CPU-only ThinkPad gave inconsistent detections. I pivoted to Roboflow's hosted inference API — same architecture, better-tuned weights, no GPU required. The .pt files still in the repo are artifacts of that failed approach.
• GPS clustering sensitivity. Initial seed data placed records too close together — the 50m clustering merged everything into one segment.

What I Learned

• Cloud inference changes what's possible without a GPU. a hosted API gave me better results with less code than local YOLOv8 ever did.
• The reframe matters as much as the code. I almost pitched "cheaper road surveys." The real product is grant eligibility — the scan output maps directly to what municipalities need to apply for STIP, MSTIP, ICIP, and FCM programs.

Built With

• cloudinary
• expo.io
• fastapi
• leaflet.js
• react
• react-native
• roboflow
• sqlite
• tailscale
• typescript