nearnet-pittsburgh

▶ Live demo: https://the-snowmen.github.io/nearnet-pittsburgh/

An end-to-end, no-backend spatial-data pipeline for Pittsburgh: it precomputes road-routed proximity facts for 115,914 buildings, publishes them as static GeoParquet/PMTiles, and lets a browser screen modeled connection scenarios without routing in the browser.

It uses public base geometry plus synthetic/proxy network attributes. The map measures distance to a modeled corridor, not a verified network; every result is a screening estimate, not a quote.

What I built

• Offline spatial pipeline: public-data ingest, EPSG:2272 measurement, multi-source Dijkstra, crossing analysis, and static GeoParquet/PMTiles export.

• Browser analytics: DuckDB-WASM recalculates the full-city screen from slider assumptions; MapLibre updates a 115k-building cost surface without a backend.

• Explorable product surface: building dossiers, routed connectors, opportunity cells, per-building KMZ export, and a mobile bottom sheet.

• Design (source of truth): docs/DESIGN.md

• Project history: docs/HISTORY.md

• POI opportunity-signal policy: docs/POI_CATEGORIES.md

What's live: the road-routed proximity screen (cost-colored building surface + reachable-within-budget sliders), a second cell-overview altitude (H3 opportunity-index choropleth that drills back into the building screen), a click-to-select building dossier (nearest modeled corridor + per-POI detail + nearest address) with per-building KMZ export, and a mobile bottom-sheet layout — all over a modeled corridor, not a verified network or operator dataset.

What it looks like

Cost surface. Every building shaded by its modeled screening estimate (green = lower, red = higher); the ramp tops out at your budget.	Building dossier. Click any building for an itemized screening estimate — routed connector distance, barrier crossings, and nearby public listings.
Cell overview. A second altitude: an H3 opportunity-index choropleth — a unitless modeled signal (not dollars) for gap-spotting.	Mobile. Responsive bottom-sheet layout — tap a building to slide up its screening estimate.

Architecture in one line

Geographic facts (distances, crossings) are baked once, offline, into static GeoParquet; cost opinions (cost-per-foot, circuity, budget) live on browser sliders. The browser does arithmetic over baked facts — no routing in the browser, ever (DESIGN.md §2).

build/ — the offline ETL (Phase 0)

Runs once on a build machine to produce the static data/*.parquet. The shipped web app ships none of this Python. The build routes/joins offline and bakes facts; see docs/DESIGN.md §2.

build/
  config.py      # locked constants + Phase-0 tunable starting guesses (single source of truth)
  sources.py     # Overture (DuckDB+httpfs) / OSM (OSMnx features + routable graph) / TIGER (pygris), clipped to the City polygon
  routing.py     # V2: real road-following connector — multi-source Dijkstra over the OSM street graph to the corridor
  geometry.py    # EPSG:2272 routed connectors, ST_Crosses-gated crossing counts, bridge proximity, POI assignment
  cells.py       # V1.5: pure GROUP BY of buildings.parquet -> H3 opportunity-index cell layers
  measure.py     # Phase-0 tunables report -> data/phase0_report.{json,md}
  emit.py        # GeoParquet (geometry back to EPSG:4326) + §9 closing-query validation
  export_web.py  # split GeoParquet -> browser assets (facts parquet, connectors, GeoJSON, PMTiles)
  precompute.py  # CLI orchestrator

Setup (conda — ETL only)

The GIS stack (geopandas/shapely/osmnx/pygris) is for the build step only. The app needs no Python.

conda create -n nearnet python=3.12 -y
conda activate nearnet
pip install -r requirements.txt

Run

# Fast end-to-end smoke test — confluence region (Golden Triangle + North Shore + South Side)
python -m build.precompute --sample 500

# Full City of Pittsburgh run (all buildings; multi-GB Overture scan)
python -m build.precompute --full

# Reuse the cached source layers from a prior run (fully offline)
python -m build.precompute --sample 500 --skip-fetch

# Geometry + measurement only, skip GeoParquet emission
python -m build.precompute --measure-only

Outputs land in data/ (gitignored, fully reproducible): buildings.parquet (one row per candidate, all DESIGN.md §9 columns) + companion layers (network, barriers_*, bridges)

• phase0_report.{json,md} (the distance/POI/crossing distributions that lock the deferred tunables D_max, POI snap distance, and corridor density).

Stack

• App: MapLibre GL JS + React/TypeScript, DuckDB-WASM over static GeoParquet, PMTiles building tiles (raster CDN basemap), GitHub Pages.
• Build: Python — Overture/OSM/TIGER ingest, EPSG:2272 geometry, DuckDB emit.

Licensing

Code: MIT. Data carries its own terms (OSM ODbL attribution + share-alike; Overture, USGS NHD, Census TIGER attribution) — see docs/DESIGN.md §12.

GitHub presentation

Use cost.webp as the repository social-preview image in GitHub’s Settings → General → Social preview. The project does not synthesize a separate marketing image: links should show the real application.