How we built AI replies with a “typing” effect
In AI CareerBot, we implemented SSE/WebSocket streaming so the AI response arrives token by token, as if the bot is typing. That immediately shows progress and noticeably increases engagement.
In AI CareerBot, we implemented SSE/WebSocket streaming: the AI response doesn’t arrive all at once, but token by token, as if the bot were actually typing — yes, the same UX everyone got used to with ChatGPT. The user doesn’t stare at a blank screen; they see progress right away. Small thing? In practice, it’s a strong engagement boost.
But at the viral launch, the classic IT amusement park ride began: thousands of concurrent users, and we had to avoid taking down the GigaChat API. What we did: added a request queue, rate limiting, and graceful degradation under load. In other words, at peak traffic the service doesn’t “die красиво,” it keeps working predictably.
The result: the interface feels alive, and the backend doesn’t go into tilt when the product suddenly takes off. In the era of AI products, that’s no longer a feature — it’s hygiene.