Context
HKUST students face interview-based major allocation, yet many lack access to relevant interview experience. General voice AI tools do not include past questions and often feel unfocused for this scenario.
Product Design
I designed and shipped modules for personalized question generation, multi-turn real-time follow-ups, and four-dimensional evaluation, using past question banks and user resumes as the basis for interview generation.
Iteration and Growth
- After the first launch, interviews had a low completion rate. User interviews and retention funnel analysis revealed slow first-question generation and unstable follow-up outputs.
- I redesigned the flow with asynchronous question generation, reducing first-question waiting time by 80%, and separated text generation from voice playback by producing structured JSON first.
- Within seven days of launch, the site reached 2,000+ visits and converted over 30% of the target audience into registered users. Users completed 200+ mock interviews totaling more than 11 hours.