TOPIK AI Writing MVP — Seol Sunwoo (Jake)

Problem: The AI feedback feature already existed — as an add-on to an online lecture platform. The problem was that learners didn't gather around it. Digging into the Vietnamese market revealed the structure: with offline academies dominant, reading and listening can be done alone, but writing has nobody to grade it — so self-study breaks exactly there. So I changed the question, from 'should we build AI feedback' to 'how do we reconnect the broken learning loop.'
Role: I owned market research, problem definition, MVP feature specs, IA/UI/UX documents, AI feature and prompt design, screen planning, and prototyping — holding everything until just before production, then handing off to the development team to raise polish. I defined my job not as delivering documents but as making the team move without interpretation cost.
Process: My first hypothesis was simple — 'accurate feedback plus accurate repetition always raises scores.' Interviews and completion/drop-off data broke its hidden premise: repetition requires learners to come back, and they were drifting away. Two forks, judged separately. Build long-form type 54 first? Good for exam prep, but long-form on mobile invites churn — I narrowed to mobile-friendly 51–52. Could UX alone manufacture motivation? It couldn't — so a binding learning incentive and a completion refund were layered into the design to fill motivation from outside the product too.
Result: The service reached closed beta (CBT); after the refund launched, learners' average goal attainment and profitability improved. The most instructive failure is also on record — content and development ran on schedule while decision alignment lagged and delayed the timeline. Ever since, I record the basis for decisions — data and judgment criteria — before documents. The next project is that answer.