Most apps for couples feel like either a chat client with streaks bolted on, or a $10/month subscription wall. I originally built quzora as a private, animated home for my own relationship: a wooden railway timeline of memories, ruled notebook pages you write on directly, a fairy-lit polaroid photo wall, wax-sealed letters, an 82-word feelings wheel, a shared list, and celebration reminders.
Once it grew into a real multi-tenant app, I wanted to keep the architecture simple enough for one person to maintain and cheap enough to stay free. Here are four technical choices that paid off the most.
1. One container, zero idle-wake pain, $4/month
Instead of splitting frontend, API, workers, and a media proxy across five services, everything ships as one multi-stage Docker container on Fly.io (React 19 + Vite + Tailwind v4 on the client, Express 4 + Drizzle ORM on the server, Neon Postgres for data, Cloudflare R2 for media).
Two small details made it feel fast without costing much:
- One always-on machine (
min_machines_running = 1): Auto-stopping saved a couple dollars a month, but a 2-second cold boot feels awful when someone opens the app just to read a letter. - A database-free
/healthzprobe: Fly polls/healthzevery 30 seconds. Because that route deliberately never touches Postgres, Neon still scales to zero overnight while the Express process stays warm.
Total monthly cost at couple scale: ~$4 for Fly, $0 on Neon, $0 on R2 (zero egress), $0 on WorkOS AuthKit, $0 on Resend.
2. Lazy WebP variants and locked-down /media/*
Photos and videos are 99% of the bytes in a memory app, and an unguessable S3 URL is not real privacy. Every media request goes through GET /media/<key> on Express, which enforces both authentication and couple ownership before streaming a byte.
To keep the photo wall snappy on phones:
- Originals are stored untouched in R2 as masters.
- Feed thumbnails (
?w=800) and lightbox views (?w=1600) are generated lazily on first request withsharp(EXIF-rotated, resize-first, WebP quality 80) and cached beside the original withCache-Control: private, max-age=31536000, immutable. A 5 MB phone photo drops to ~90 KB. - Videos never download in the feed.
ffmpegextracts a single poster frame lazily (?poster), and tapping play streams the original via HTTP range requests.
3. Importing from Google Photos (and the COOP popup gotcha)
Google deprecated full library access in favor of the Google Photos Picker API, which is actually better for privacy because the app only ever sees the exact photos the user selects.
The sneaky bug: Google's picker sends Cross-Origin-Opener-Policy, which severs the popup from window.opener. That means postMessage fails and popup.closed reads true while the picker is still wide open.
The fix was treating the popup as fire-and-forget and having the main app tab poll a lightweight status endpoint (connecting -> picking -> ready). Once ready, the server streams the import back as NDJSON so each polaroid lands on the photo wall the moment its WebP variant is ready, carrying its original createTime date.
4. Prerendered public pages + Markdown twins without Next.js
Because the signed-in app is a heavy client-side SPA, I didn't want to rewrite everything in a meta-framework just for SEO. Instead, npm run build runs a 60ms Vite SSR pass (renderToString) over the public landing page, guides, and a set of free zero-login couples tools (a days-together calculator, an interactive feelings wheel, an open-when prompt picker, and a bucket list generator).
During that same build step, a small script converts each prerendered HTML page into a clean .md twin (stripping SVGs, nav chrome, and decorative mockups). Express serves the Markdown version whenever an AI agent or curl asks for .md or sends Accept: text/markdown, alongside llms.txt and JSON-LD schemas.
If you'd like to try the app, it's free to use. Check it out at quzora.com.