Type
City guide · SvelteKit
A
Angkor Now
Case study · Live demonstrator
A living guide to Siem Reap, run by a machine.
Angkor Now is a tourism & culture guide for Cambodia’s temple town: events, venues, festivals, maps and guides. Behind its quiet editorial surface sits an autonomous ingestion pipeline that reads the web, extracts structure with an LLM, de-duplicates, and stages everything for a human before it ever goes live.
176 venues mapped
AI-ingested · human-approved
The brief
One town, a hundred sources, no staff.
Siem Reap’s nightlife, events and venue scene lives scattered across Facebook pages, partner listings and word-of-mouth. No single source is complete, none are structured, and most are out of date within a week.
Angkor Now’s premise: a city guide that maintains itself. The site continuously reads those messy sources, uses a language model to pull clean structured records out of them, and routes everything through automated de-duplication and a single human approval gate before publishing, so the public site stays trustworthy while the grind stays automated.
The result is a fast, warm, editorial product that reads like it has a full newsroom behind it, built and sponsored by AI Boffin Hub as a functional demonstrator of exactly that idea.
“The scraper never touches the database. It writes JSON. A human decides what becomes real.”
176
Venues mapped & geocoded
100%
Pages prerendered & static
2
Dedup algorithms · exact + fuzzy
0
DB writes from the scraper
1
Human review gate before publish
The product
A guide that feels hand-edited.
Cormorant Garamond headlines, warm paper, a forest-and-turmeric palette and theme switching: every surface is editorial first, database second. A tour of what a visitor actually sees.
01 / What's on now
“What’s on tonight.”
The default question a traveller asks, answered first. A timezone-correct timeline of tonight’s events, all-day then by start time, with category and price tags inline, rolling straight into “coming up next” for tomorrow.
02 / The events feed
Every event, filterable.
The full directory of upcoming events with faceted filtering by category, date and venue. Each record is an AI-extracted, human-approved entry: recurring nights, ticketed shows and free meetups all normalised into one schema.
03 / Plan ahead
A month at a glance.
The same event corpus, pivoted into a monthly calendar for trip planning. Days light up with what’s happening; click through to the day or the event. Built from the identical prerendered data, with no second source of truth.
04 / The directory
Bars, restaurants, culture.
A browsable venue directory across bars, restaurants, cafés, music and entertainment venues. Each has its own prerendered detail page, opening hours, neighbourhood and a generated QR code for sharing on the ground.
05 / Find a venue
An interactive map, clustered.
All 176 venues plotted on a MapLibre GL map over CARTO tiles, with marker clustering, category-coloured pins, an “Open Now” filter and a neighbourhoods overlay. The side rail and map stay in sync as you filter and search.
06 / Editorial guides
Long-form, curated.
Hand-shaped guides (the best rooftop bars, apsara dance, getting around, family-friendly days, healthcare & emergencies) that wrap the structured data in human context. This is where the machine-run guide still reads like a person wrote it.
07 / Find anything
“Find anything.”
One search box across venues, events and guides, with suggestion chips for common intents: cocktail bar, apsara dance, kandal village, festival. Below it, a persistent emergency band keeps the Royal Angkor Hospital line one glance away.
Under the surface
Advanced features, quietly done.
The polish a traveller never notices but always benefits from: correctness, speed and resilience baked into the architecture.
Timezone-correct everything
All dates normalise to Asia/Phnom_Penh at intake, so “tonight” means tonight in Siem Reap regardless of where the visitor or the server sits.
Dual-algorithm dedup
An exact key match catches re-scrapes; a fuzzy title-date-venue similarity score catches the same event described two different ways across sources.
Marker clustering at scale
MapLibre GL clusters hundreds of geocoded pins into readable groups, recolouring and recounting live as you zoom, pan, filter by category or toggle open-now.
Fully prerendered
SvelteKit prerenders every route to static HTML on deploy: instant first paint, no cold database calls in the request path, trivially cacheable on a CDN.
Theme & dark mode
A three-accent theme system (forest, turmeric, oxblood) plus light/dark, all driven by CSS custom properties set in one token file. No hardcoded brand colours anywhere.
QR & sharing built in
Venue pages generate QR codes server-side for on-the-ground sharing, with clean Open Graph metadata so any link unfurls into a proper card.
The AI backend
How the guide reads the web for itself.
A seven-stage intake pipeline turns messy public sources into clean, de-duplicated, human-approved records. Every stage is isolated, and the one irreversible action, publishing, always needs a person.
01
Source fetch
fetch-source.ts · pluggable fetchers
Each source (partner listings, Facebook pages via Apify actors, curated feeds) has its own fetcher returning a common
RawEvent[] shape. New sources plug in without touching the rest of the pipeline.02
Read-only scrape
scrape-events.ts · JSON out, zero writes
The scraper is deliberately powerless: it only writes a JSON file and never touches the database. Scraping and persistence are fully decoupled, so a bad scrape can never corrupt live data.🔒 Hard isolation boundary
03
LLM extraction
llm.ts · OpenRouter
Raw, unstructured text and HTML is passed to a language model via OpenRouter, which returns structured fields (title, start & end time, venue, price, category and tags) out of prose a regex never could.
04
Normalise & filter
normalise.ts · timezone · blocklist
Extracted records are coerced into the canonical events schema: times shifted to
Asia/Phnom_Penh, slugs generated, categories mapped, and a blocklist drops spam and non-events before they go further.05
Dual-algorithm dedup
dedup-check.ts · exact + fuzzy
Two layers decide each record’s fate: an exact match on source identity, then a fuzzy similarity score across title, date and venue. The output is a decision:
insert, merge or stage.06
Human review gate
admin/pending · staging
Anything uncertain lands in a staging queue where an editor approves, merges or rejects it. Nothing the machine is unsure about reaches the public site unsupervised.✋ Required human approval
07
Enrich, persist & ship
Apify · Supabase · Netlify
Approved records are enriched (images via Apify, venue geocoding), written to Supabase Postgres, then the site re-prerenders and deploys to Netlify’s CDN, back to a fast static guide.
scripts/intake/run-intake.ts — the orchestrator
// fetch → normalise → dedup-check → (insert | merge | stage)
const raw = await fetchSource(source); // RawEvent[]
const events = raw.map(normalise); // → events schema, TZ-correct
const clean = events.filter(notBlocked); // blocklist
for (const ev of clean) {
const verdict = await dedupCheck(ev); // exact + fuzzy
if (verdict.kind === 'duplicate') skip(ev);
else if (verdict.kind === 'merge') mergeInto(verdict.target, ev);
else stageForReview(ev); // ← human gate
}
// the scraper itself? it only ever wrote a .json file.Isolated by design
Scrape, extract, dedup and publish are separate stages with hard boundaries, so failures stay contained instead of cascading into live data.
Human-in-the-loop
The model proposes; a person disposes. Automation handles the volume; the single approval gate handles the trust.
Model-agnostic
Extraction runs through OpenRouter, so the underlying model can be swapped without rewriting the pipeline around it.
The stack
Built to stay cheap and fast.
A deliberately lean, mostly-static architecture: sophisticated where it counts, boring everywhere else.
SvelteKit
Framework · SSG
Every route prerendered to static HTML for instant loads.
Supabase
Database
Postgres of record for approved events, venues and guides.
Netlify
Hosting · CDN
Static deploys served globally, rebuilt on new data.
OpenRouter
LLM gateway
Model-agnostic extraction of structure from raw text.
MapLibre GL
Maps
Clustered, filterable venue map over CARTO tiles.
Apify
Scraping · assets
Facebook actors and image/asset enrichment.
TypeScript
Language
End-to-end types from intake scripts to UI.
tsx scripts
Intake jobs
The pipeline runs as standalone, schedulable scripts.
Why it holds up
Autonomous, but never unaccountable.
JSON-only
The scraper can’t write to the database; it physically only emits files.
Staged
Uncertain records wait in a review queue, never auto-published.
Sourced
A public “Our sources” page documents where data comes from.
Static
Prerendered output means no live DB in the request path to fail.
A functional demonstrator
See it running.
Angkor Now is live: a real, self-maintaining city guide built and sponsored by AI Boffin Hub to show what an LLM-run editorial product looks like in production.