Conversation platformLive2026
Chat Translate
Korean–Vietnamese messaging, with the translation under every message.
The problem
Korean and Vietnamese staff share a work group, a sales thread, a support shift. Translating means switching to another app and pasting the reply back, which breaks the conversation. Raw machine output also gets the register wrong: it addresses a customer too casually, or answers a manager in clipped fragments.
The solution
Every message shows twice, the original and its translation. Each room sets a register (work, friends or intimate), and each member is given a form of address, so the translation keeps the right level of respect instead of flattening it. The admin console tracks failure rate, latency and cost per model, and does not display message content.
The screens
- 01Two lines per bubble: the original and its translation
- 02The room’s register, and how to address each person in it
- 03Conversations with their latest message
- 04Three plans, sized by translated messages per day
- 05AI operations: success rate, p50/p95 latency, cost by language pair
Key features
Translation in the thread
Original and translation share one bubble, so nobody leaves the app.
Register per room
Three language registers; a change reaches every live conversation within 30 seconds.
Slang dictionary
A term’s meaning is injected into the prompt whenever a message contains it.
Forms of address per person
Set the relationship to each member, and the translation addresses them in the right register.
Quality measured by votes
Readers rate translations; the QA screen groups them by model, cause and language pair.
Private by design
The admin console sees metadata only; message content is never shown here.
Stack & numbers
- 96,6%
- Translation success rate
- 937ms
- p50 response
- 1,6s
- p95 response
- NestJS
- PostgreSQL
- Redis
- Socket.IO
- OpenAI
- React Native
- Zalo Mini App
- AWS