YEAPP was KRAFTON India’s voice-first messenger: you talked instead of typing, and Ziggy, an AI companion, talked back. It was already on the Play Store when I joined in late July. In a new user’s first minutes, nobody they knew was on the app yet, and the tour that taught it was built into the app, so changing a line or a voice meant shipping a new version of the app.
A voice tour that changed without an app release
Ziggy taught a new user the app in a chat, by voice, with captions that lit up word by word. I moved the tour’s script and voice clips from the app to the backend. Queued jobs synthesized each of the 58 clips, transcoded it, aligned every word with Whisper and published the audio to a CDN under a content hash. An admin could edit a line or upload a recording by a person instead, which ran the same pipeline without the speech synthesis. A kill switch could turn the tour off without an app release.
Messages that wait for a friend to join
The app listed contacts who weren’t on YEAPP yet and let you message them. The server held the messages, within a limit and an expiry set from the admin console, and delivered them when the contact finished signing up. If the first message was a voice note, it went into the invite video the sender shared.
A backup vendor that took over
The companion had a fallback chain in its config, but it never took over. A 30-second breaker around the whole chain gave up before the second vendor was tried, and the vendor adapters returned an error value instead of throwing, so their circuits never opened. I fixed both, set up a second vendor in case the first ran out of credit, and tested the failure path live.
Stickers
Sticker search returned results as you typed, from a Redis cache that every write cleared. I also built two sticker pipelines: auto-tagging stock stickers with a vision model, and turning sticker videos into animated WebP with a transparent background.