Lilybee

Lilybee animates a child's crayon drawing and plays it back in augmented reality on the original piece of paper. A parent photographs a drawing they already have, optionally records a few seconds of their kid's voice, and gets back a short film that lives on the fridge — point a phone at the drawing and it moves and talks. Nothing is printed or shipped: the artwork already on the wall is the marker.

Behind that one tap is a full generative pipeline. The order is the single source of truth, and a Firebase state machine walks it through every stage — signed-URL upload and a perspective-warp crop, Whisper transcription of the voice clip, prompt synthesis with Claude, a sound bed from ElevenLabs when there's no recording, video generation with Kling, then an ffmpeg/ImageMagick service on Cloud Run that packs the result into the alpha-stacked video the app plays back. Stripe sits in the middle, which means a failure anywhere has to be recoverable rather than merely logged.

AR playback runs on expo-ar-marker — the same image-tracking module I wrote for InkStory, now on its second product. That's the part I'd point at: I built it once because the platform didn't offer it, and it paid for itself the second time. The rest is mine end to end too — the Expo app, the ordering site, the Cloud Functions backend, the render service, and the deploy pipelines under all four.