Curio Camera

Point your camera and ask out loud — AI answers by voice, and even lights up the answer right on your photo. Snap a floor map, ask "Where's the restroom?", and see it glow. Works in 15 languages.

Comment

Inspiration

There are things you wonder about but never look up — a flower on a walk, a statue in a park, a floor map at the mall. And there are moments when you literally can't look things up: your hands are full, you're walking, your kid is asking "What's that?" for the tenth time. I wanted an app where the camera is the question and your voice is the keyboard — and the answer comes back out loud, so reading is optional.

What it does

Curio Camera ("ものしりカメラ" in Japan): point, snap, ask out loud.

  • Ask by voice or text about anything you photograph — AI answers in text and reads it aloud
  • The answer can light up on the photo itself. Snap a floor map, ask "Where's the restroom?", and the exact spot glows
  • Ask "Where am I?" and it combines your location with nearby place candidates, answering with the facility name and a map card
  • Follow-up questions keep the context of the photo
  • Works in 15 languages — ask in yours, it answers in yours
  • Question history (last 100), with iCloud backup as a premium perk

How I built it

  • SwiftUI, deliberately one screen, iOS/iPadOS 18–26, iPhone & iPad
  • No API key ships in the app. All AI calls go through a Cloud Functions proxy (askClaude) that streams Claude API responses over SSE. Firebase anonymous Auth + App Check verify every request, and rate limits / premium checks live server-side
  • RevenueCat powers the subscription — premium (unlimited questions, high-res vision, iCloud backup) is gated purely by the premium entitlement, never by a local flag
  • CloudKit backup with a merge-only restore that never overwrites local history, even if the cloud returns zero records
  • Japanese TTS loves to misread kanji, so speech goes through a 2-layer fix: the model emits pronunciation hints that swap only the audio (the display keeps the kanji), plus a curated reading dictionary applied at the single speech exit point

Challenges

  • One App Review rejection taught me two lessons. I was showing error.localizedDescription raw on screen — now every error goes through a single user-facing presenter. And App Check takes different paths in Debug (debug token) vs Release (App Attest); all my device testing ran on Debug, so the missing App Attest registration only surfaced at review. Now every #if DEBUG branch gets verified on both sides via TestFlight
  • A Release-only optimizer crash (EarlyPerfInliner + a generic UIHostingController held by a Coordinator) blocked archiving while Debug built fine

What's next

A use case I didn't design for emerged from real post-launch feedback: parents out with kids. Floor maps, nursing rooms, endless "why?" questions — exactly the moments when you never have a free hand. I'm building toward that, and sharing the whole journey on X with #Shipaton #BuildInPublic.

Built With

Share this project:

Updates

Submission history