BedRocked

Sequencing Somerville's $1.29 B sewer-separation program from data instead of by hand.

🔗 Live demo → https://sewershed-bedrocked.vercel.app

Built at the Physical-AI Hackathon (Somerville, MA · June 2026) — sponsored by Cyvl · Autodesk · NVIDIA · City of Boston.

---

The problem

Somerville, MA is under a $1.29 billion, multi-decade mandate to separate its combined sewers (pipes that carry sewage and stormwater together and overflow into the Mystic and Charles during storms). 2,404 street segments are still combined. Today the city decides which streets to dig, and in what order, largely by hand — across pavement condition, pipe age, dig depth, nearby assets, and water-main risk.

BedRocked turns that into a data product: it scores every combined segment 0–100 for dig-readiness and lets a planner sequence the whole program on a map.

What it does

• 🗺️ Interactive map of all 2,404 combined-sewer segments, colored by a weighted 0–100 readiness score (red = dig first → green = later).
• 🧮 6-factor scoring model — pavement urgency, pipe age, dig depth, asset bundling, network leverage, and water-main co-risk — joining city sewer GIS to Cyvl's street scan.
• 🧰 Layer toggles — road condition (PCI), water mains & risk, catchments, storm drains, catch-basin condition, heatmap.
• 🔎 Plain-English search — "high-priority pipes older than 100 years" — a Claude model turns language into a deterministic filter (it never invents numbers; scoped to this dataset only).
• 🖼️ Evidence cards — click any segment for its score breakdown, pipe facts, and the live Cyvl street-scan photo as provenance.
• 📐 Autodesk export — one-click DXF dig-plan (layered by priority) that opens in AutoCAD / AutoCAD Web, plus an APS Viewer pipeline that renders the plan in 3D.

Data & provenance

Two independent, real sources — nothing mocked. Full sheet: docs/Sewershed_Data_Sources.pdf

From Cyvl (AI street scan)	From Somerville GIS (public)
Pavement condition — 5,080 segments	Sewer mains — 11 layers, 2,404 combined
Catch basins — 381 (+ photos)	Catchments — 7 drainage areas
Above-ground assets — 8,254	Storm inlets — 3,659
Signs 3,782 · markings 7,116	Water mains 8,028 · risk 2,061
Street-level imagery (live CDN)	Pipe age / depth / material

Evidence-card photos stream live from Cyvl's CDN — verifiable provenance, not a label.

The catch-basin condition model (knowledge distillation)

Every Cyvl asset ships with condition: null. To fill it for the 381 catch basins:

Cyvl photo → crop to the projected asset marker
           → Claude Vision labels 112 crops  (teacher)
           → distilled classifier scores all 381  (student)

Honest framing: at 381 basins this proves the scaling pattern — the teacher labels once, the lightweight student then scores every asset for cents, city-wide.

Architecture

                 ┌─────────── Cyvl API / GIS exports ───────────┐
                 │  pavement · assets · imagery · sewer · water  │
                 └───────────────────┬──────────────────────────┘
        spatial join + 6-factor scoring (Python / shapely)
                                     │
                         segments_scored.json  ◀── one precomputed file
                                     │
        ┌────────────────────────────┴───────────────────────────┐
   React + MapLibre map            serverless /api/search (Claude, scoped)
   (Vite, deployed on Vercel)      Autodesk: DXF export + APS Viewer

Stack: React · Vite · MapLibre GL · Python (shapely, scikit-learn) · Anthropic Claude (NL search + vision labeling) · Autodesk Platform Services · Vercel.

Run locally

# backend (scoring API)
cd backend && pip install -r requirements.txt && uvicorn main:app --port 8000

# frontend
cd frontend && npm install && npm run dev      # → http://localhost:5173

The deployed build serves the data statically (no backend) and runs NL search as a Vercel serverless function — keys live only in Vercel environment variables.

Repo

Path	What
frontend/	React + MapLibre app (deployed)
backend/	FastAPI scorer + endpoints
autodesk/	DXF export + APS Viewer pipeline
cyvl_data/ · data/	Cyvl exports + Somerville sewer GIS
docs/	data-provenance sheet

---

Team

Built collaboratively by our team at the hackathon — the shared repo we worked in: https://github.com/yashphalle/Cyvl-Hack

Sponsors: Cyvl · Autodesk · NVIDIA · City of Boston Physical-AI Hackathon · Somerville, MA · 2026