RoadWatch
RoadWatch is an autonomous multi-agent pavement surveyor that grades road condition at scale using Mapillary street-level imagery — no custom CV model, no windshield survey crew, no $250k engineering contract.
Core trick
Grades pavement from Mapillary street imagery and uses SHA-256 hashing to catch backdated listings that are really duplicate images.
Description
RoadWatch is an autonomous multi-agent pavement surveyor that grades road condition at scale using Mapillary street-level imagery — no custom CV model, no windshield survey crew, no $250k engineering contract.
A 3-tier Opus 4.7 hierarchy works like a field crew:
Street Captain (1×) plans the inspection route, dispatches survey teams, synthesizes a corridor narrative
Point Surveyors (up to 3×, parallel) investigate each survey point across all available years — cross-witnessing temporal change
Year Investigators (up to 2× per Surveyor, parallel) examine the same spot from different epochs, sharing claims on a blackboard so siblings can anchor their yaw angles
Peak concurrency: 6 simultaneous Opus 4.7 calls. The discipline gate is deterministic — it inspects the visit log and refuses a grade() call if multi-year coverage exists but only one year was examined.
A SHA-256 collision detector identifies when "multi-year" listings are actually the same image with backdated metadata — surfacing a data integrity problem invisible to any pipeline that trusts timestamps.
The web UI lets a public-works engineer draw a polygon, select streets, watch the agent hierarchy work in real time — cursors walk the block, evidence cards stream in, the map recolors as grades land. One click, ~$5, any street in the world with Mapillary coverage.
Validated: Spring St LA, Chestnut Ventura, SoHo Broadway. Calibrated against LA's published 69,282-segment PCI dataset.
README excerpt
# RoadWatch
> Agentic pavement-condition triage for DOT engineers — a 3-tier multi-agent
> Opus 4.7 fleet that reads street-level Mapillary imagery, grades every
> survey point on a corridor, and turns the evidence into a decision-grade
> priority brief.
🎬 **[3-minute demo video](https://youtu.be/XVfxFR1nMwg)**
🌐 **[Live demo](https://bibek624.github.io/roadwatch/)** · **[Source on GitHub](https://github.com/bibek624/roadwatch)**
Built for the **Anthropic Opus 4.7 hackathon** (April 2026).
---
## What it is
Public works departments need to answer one question every Monday morning:
**which streets need attention first, and what's wrong with them?** Today they
either wait for a $250k windshield survey every 3 years or guess from 311
complaints. The visible answer — the cracks, the patches, the alligator
fatigue, the failed utility cuts — is already in Mapillary's open street-level
imagery. Nobody could read it at scale before, because pavement-condition
grading needs an engineer's eye.
RoadWatch deploys a **3-tier hierarchical Opus 4.7 agent fleet** at every
survey point:
```
┌──────────────────┐
│ Street Captain │ ← plans, dispatches, synthesizes
└────────┬─────────┘
│ dispatches up to 3 in parallel
┌────────────┼────────────┐
▼ ▼ ▼
┌──────────┐ ┌──────────┐ ┌──────────┐
│Surveyor 1│ │Surveyor 2│ │Surveyor 3│ ← one per survey point
└────┬─────┘ └────┬─────┘ └────┬─