Case study · 2025–2026
A collaborative mapping platform where creators build private maps and earn money from them. Designed, built and shipped by one engineer.
Joshua Wong — solo founder-engineer · product, backend, frontend, mobile, infra
The map behind this page is live. Scroll to move it.
Try it
This is the actual Allmap frontend running on a frozen snapshot of its data: every public map, place, trail and event. Browse, search and open anything; creating and editing are switched off.
01 · The problem
Travel creators, guides and local experts pour knowledge into Google lists and spreadsheets that they can't brand, gate or charge for. Allmap gives them a map of their own: private or public, collaborative, with paid access built in.
How it works · Layers
Every map is a layer: restaurants, climbing crags, pickleball courts, hot springs. Stack the ones you like and hide the rest. Click a pin on the map.
How it works · Pins
A restaurant, a viewpoint, a friend's house. Each pin carries photos, an address, tags and links, and opens with directions one tap away.
How it works · Events
Concerts, markets, meetups. Ticketmaster events stream in automatically, refreshed every day across 249 city cells.
How it works · Routes
Draw it point by point, upload a GPX file or record it live. Distance, elevation and the profile are worked out for you. The amber lines are 1.1M real trails.
How it works · Messages
A tip, a label, a comment for whoever comes by. Messages sit on the map as speech bubbles.
How it works · Paint
Turn Paint on and color the shared map by category. Every cell painted earns karma.
How it works · Map details
Open any map to see its locations, feed, messages, links and members. Click an item to fly straight to it.
02 · What I built
Django 5 + DRFPostgreSQL + PostGISNext.js 16MapLibre GLCapacitor 8Stripe ConnectCloudflare R2Render
03 · Architecture
Every map is a stack of layers: pins, routes, drawings, paint grids, text. Each layer type is one self-contained handler that owns its own data format, validation and updates. The API finds handlers on its own, so adding a layer type means writing one backend class and one frontend component.
Why it matters: five layer types shipped without touching the core. The registry stays thin and each handler stays deep.
04 · Scale & performance
Problem: streaming route GeoJSON from the API stalled at scale.
Decision: pre-render routes into PMTiles vector tiles with a parallel tippecanoe pipeline (6× faster), hosted on Cloudflare R2.
Result: a 2.1 GB tileset served as byte-range requests from a CDN, with no API server in the path. That's what you're seeing on this map.
Problem: the map list fired an N+1 query storm.
Decision: prefetching, dynamic field selection and slim marker types at ingestion.
Result: 93% fewer queries per load. Clustering moved to a web worker, and the cache uses LRU eviction.
Problem: Ticketmaster caps calls per day and per second.
Decision: 10 staggered cron jobs across 249 city cells in tiers, ordered so polling always reads fresh counts.
Result: about 4,500 calls a day, always under 5 req/s, with a 500/day on-demand safety valve.
05 · AI-first engineering
I build with AI coding agents as the primary implementers and act as architect and reviewer. Speed alone doesn't make that work. Discipline does, so I wrote it down:
CLAUDE.md): deep modules, information hiding, errors defined out of existence. Agents follow it on every change.06 · Results & next
AI agents make writing code cheap. They make design the bottleneck. The most valuable thing I built isn't a feature. It's the set of rules that keeps the codebase simple while it grows fast.