AInterviewMock Interview Service

Mock Interview Service

AInterview

Background

Research showed 46% of job seekers find interviews the hardest part of job hunting. The project was born from recognizing a critical lack of opportunities to practice in a realistic, repeatable environment.
- Existing prep services only provided question lists. No service combined real-time voice interaction, role-specific questions, and AI feedback.
The Likelion 12th cohort final project — in a 4-person team, I handled planning, design, and frontend development over about 4 months (September–December 2024).

Analyzed actual pain points of job seekers and defined three core features: voice-based interview simulation, auto-generation of role-specific questions, and AI feedback — then prioritized them.
- Implemented speech-to-text (STT) and text-to-speech (TTS) conversion using Azure Speech Service, enabling voice-based interaction similar to a real interview.
- Fine-tuned GPT-4o-mini on role-specific interview data to design a system that auto-generates appropriate interview questions for IT, business, marketing, and other fields.
Designed the complete user journey: Login → Role Selection → AI Question Generation → Voice Answer → STT Conversion → Feedback.

Designed an interactive UI replicating a real interview environment in Figma and implemented it in React. Included elements to enhance interview immersion: voice recording visualization, timer, and question cards.
- Integrated with Spring Boot backend API to implement real-time STT processing, question generation, and feedback retrieval flow.

Completed the voice-based AI interview simulation service and presented it as the Likelion 12th cohort's final project.
Delivered a service capable of generating role-specific interview questions and voice-based interaction through the combination of Azure Speech and GPT fine-tuning.

STT took seconds to transcribe each answer — time we couldn't shrink with tech. So we filled the silence with a timer and progress cues. This was my first practice in accepting a technical limit and wrapping it in UX.
We kept negotiating the GPT fine-tuning scope with the dev team. Cutting scope between 'ideal question generation' and 'what ships on time' — that was the real substance of the PM role.