Open to Summer 2027 · data / analytics / AI engineeringList view

Encore — Live-music app

Beta×

A live-music app where an AI pipeline reads venue websites every day and decides which shows are reliable enough to publish.

221 venues · 2 cities
Atlanta and South Bend venues, pre-launch beta
  • 44Postgres migrations
  • 32tables, all with row-level security
  • 4scheduled jobs (ingest, digests, reminders)
  • 82pages across fans, musicians, venues, admin
Encore's Tonight feed for Atlanta: genre filter chips above cards for Sweet Georgia's Juke Joint, Aisle 5, Commune ATL, Boggs Social & Supply, 529 and Buckhead Theatre, each with a time, price and distance.
The Tonight feed: every show the pipeline published, soonest first.
Encore on a phone: the Atlanta feed with a search bar, filter chips, show cards and a bottom tab bar.
The same feed on a phone.
Encore's map of metro Atlanta with green markers for shows tonight and orange markers for upcoming shows, clustered around Midtown and Little Five Points.
Map view: green pins are tonight, orange are upcoming.

Open the live app ↗

The problem

Live music in Atlanta is spread across hundreds of venue websites, Instagram posts and calendars, each in a different format, and small rooms like bars and breweries never show up anywhere. Finding a show tonight means checking a dozen sites, so Encore pulls it all into one feed and map — and the hard part is the data.

What I built

  • AI ingestion with confidence gating. A daily job scrapes each venue's events page, and an LLM extracts events into a strict Zod schema with a confidence score. Under 0.55 is dropped; 0.8+ from venue-verified sources (0.9+ for admin-added ones) publishes automatically; everything in between waits in an admin review queue.
  • Messy-data cleanup. Free-text performer names are matched to artist profiles with Postgres trigram similarity (0.7+ is a match, lower creates a stub flagged for review). User-submitted venues are normalized by an LLM, geocoded and de-duplicated before an admin approves them.
  • Source health monitoring. Every venue source runs in isolation, and a dashboard buckets each one as healthy, stale, failing or never-run from its success rate over the last 10 runs.
  • Postgres + PostGIS backend. 44 migrations, 32 tables with row-level security on every one (112 policies), 52 SQL functions and PostGIS radius search behind the Tonight feed.
  • A full product for four audiences. Fans get the Tonight feed, map, search, RSVPs, follows and groups; musicians post gigs; venues manage their own sources; admins run the console. A weekly digest email and SEO landing pages are built for launch.

How it works

  1. Venue sites
  2. Daily scrape
  3. LLM → typed events + confidence
  4. Gate: drop · review · publish
  5. Postgres / PostGIS
  6. Tonight feed + map

Stack

  • TypeScript
  • Next.js 14
  • React
  • Supabase
  • Postgres
  • PostGIS
  • Vercel Cron
  • Vercel AI SDK
  • Firecrawl
  • Mapbox GL
  • Zod
  • Resend
  • Turborepo
  • GitHub Actions

Shows

Proof

Commits84, solo, May 29 – Aug 18 2026git history
Database44 migrations · 32 tables · 112 RLS policies · 52 SQL functions · 101 indexescounted from the migration files
AI thresholdsdrop < 0.55 · auto-publish ≥ 0.8 / 0.9 · artist match ≥ 0.7ingestion and artist-matching code
Coverage221 venues seeded (148 Atlanta, 73 South Bend)seed migrations
CI/CD4 GitHub Actions workflows: checks, PR previews, production deploys, mobile build.github/workflows

What I'd do next

Measure extraction accuracy with a labeled set of venue pages, and launch the web beta in Atlanta, then bring back the Expo mobile app.

May 2026 – present · Live product. The code is in a private repo, so this page shows the design and the numbers instead.

← Back to the desktop All projects