Reach — AI siting copilot for Abu Dhabi

Reach is a single-screen, full-bleed map app that visualizes how easily residents of Abu Dhabi can reach six categories of everyday essentials (a "15-minute city" analysis), surfaces the worst-served neighborhoods ("access deserts"), and uses an LLM to recommend the single best real, buildable land parcel to place a missing facility — then lets you simulate that placement and watch the access map improve in real time.

The core story: a naive baseline (nearest-distance math) picks a geometric point that may land on water or non-buildable land, whereas Reach AI picks a real parcel that is zoning-aware. The app shows both side by side ("A vs B").

Stack

• Next.js 14 (App Router), React 18, TypeScript (strict)
• Tailwind CSS 3.4
• maplibre-gl 4.5 (CARTO dark-matter basemap, no token)
• h3-js 4.1 (hex binning)
• framer-motion 11, lucide-react
• openai 4.55 SDK pointed at Groq (llama-3.3-70b-versatile), server-side only

Quickstart

npm install
npm run build-data   # downloads the Abu Dhabi CSVs (if missing) and writes public/data/reach.json
npm run dev          # http://localhost:3000

AI (optional)

The app works fully without any key via deterministic fallback siting. To enable live streaming rationale from Groq:

cp .env.example .env.local
# set GROQ_API_KEY=your_key

The key is used only server-side in app/api/ask/route.ts and app/api/recommend/route.ts.

Data

scripts/build-data.ts auto-downloads any missing CSV from:

https://huggingface.co/datasets/eVoost/abu-dhabi-ai-proptech-challenge/resolve/main/<file>.csv

Files: osm_amenities.csv, sample_listings.csv, sample_communities.csv, sample_parcels.csv, districts.csv.

How it works

1. Engine (lib/engine.ts) — pure access-scoring math (haversine, per-category distance thresholds, persona-weighted access, worst-category detection, color ramps).
2. Builder (scripts/build-data.ts) — parses CSVs, bins listings into H3 res-8 cells as a population proxy, computes nearest-amenity distances per cell, scores per persona, and writes a minified reach.json.
3. Client (lib/client.ts) — ranks access gaps, builds naive baseline points, finds candidate parcels, and simulates a placement instantly (no network).
4. App — MapLibre heatmap of H3 access cells + amenity dots, an Inspector that streams an AI rationale, and a simulate mode that tweens the affected hexes greener.