Document translation without losing accuracy
Tarjimly’s document requests often involve medical letters, legal forms, housing documents and personal identification. A delayed translation can hold up the help someone needs.
Translators were starting each document manually. FirstPass introduced an AI draft, but that draft still needed a person who could check the meaning, correct errors and understand the context.
My role from workflow audit to design QA
I audited the existing translation journey and mapped where AI could reduce work and where a translator needed to stay involved. I designed the draft, editing, warning, confirmation and approval states with product and engineering, then followed through on handoff and design QA.
Designing around uncertain AI output
The workflow had to handle unclear scans, low-quality uploads, different language directions and sensitive content. Translators needed speed, but accuracy and cultural context could not become secondary.
That shaped the main decision: present the AI output as something to review, with a clear distinction between the source, the draft and the approved translation.
Making the source and translation editable
An error in extracted text can carry into the translation. I kept the source in the workflow and made both the extracted text and translated output editable.
This gave translators a way to fix the starting point as well as the result. They could work through the document without treating the generated text as a finished answer.
Review warnings and explicit approval
The draft could look complete before anyone had checked it. I designed a warning when a translator tried to submit without making an edit, followed by an explicit final approval.
An edit alone cannot prove accuracy. The review prompt made the responsibility visible; the translator still needed to check the document and decide whether it was ready.
Testing difficult documents and submission states
We tested different document types, language directions, unclear uploads and sensitive content. We also checked attempts to skip review, rather than testing only the successful path.
During QA, I checked that warnings, review prompts and confirmation states behaved as intended. The editing and approval patterns were designed for reuse in other AI-assisted workflows.
Translation time reduced by approximately 87%
FirstPass reduced average translation time from 37 minutes 54 seconds to 4 minutes 50 seconds, while keeping review, correction and approval with the translator.
What I learned about AI-assisted review
This project reinforced why I pay attention to what happens after AI produces an answer. A faster draft is useful when people can understand it, correct it and take responsibility for the final work.
