The wait for language support
People use Tarjimly for medical care, legal help, housing support and other urgent situations. They were waiting 60–90 seconds to reach a translator, often without knowing whether help was on the way.
Some requests needed a quick translation. Others needed a person who could understand the situation. Spark needed to give useful help sooner while keeping that distinction clear.
What we found in over 10,000 chat logs
We reviewed over 10,000 chat logs. Two types of request stood out: short translations and clarifications, and sensitive situations requiring empathy, context and human judgement.
That finding shaped the product direction. Spark could help with the first part of support, while a translator handled the requests that needed more care.
Immediate AI support with access to a translator
I designed two connected paths. Simple requests could receive a response inside the chat. Sensitive or unclear requests could move to a translator, and users could ask for a person at any time.
The user needed to understand the available help without having to make sense of how the AI worked behind it.
Transferring context during human handoff
When a translator joined, the earlier conversation stayed visible. Someone asking for help should not have to explain the same situation again because the type of support changed.
I designed the confidence and review states to show when more context or human review was needed. The product needed to be useful without appearing certain about something it could not safely resolve.
Processing, confidence and review feedback
I worked through processing, confidence, review and completion feedback. These details helped explain what was happening between the request and the response.
I used Lottie for product micro-interactions and Jitter for motion videos and reusable GIF, SVG and JSON assets for product, web and marketing.
Initial wait and abandonment outcomes
Spark reduced the initial wait from 90 seconds to under 3 seconds. Abandonment fell from 25% to 8%, while translators could focus more time on cases needing human care.
What I learned about AI and human handoff
The important decision was where immediate AI support should end and a person should step in. I wanted the user to get help quickly and still have a clear path to someone who could understand the full situation.
