AI interview preparation platform
PrepLens
A real-time voice interviewer, AI résumé analysis and instant-scored mock exams that cut manual evaluation by 70%.

Solution
PrepLens helps people prepare for interviews. Its centrepiece is a real-time, speech-to-speech AI voice interviewer: you answer out loud, it responds out loud, and at the end it gives feedback, a score and suggestions.
Voice interview
- System: Candidate picks a job role
- Model output: Interviewer asks a question
- Human decision: Candidate answers aloud
- Model output: Answer transcribed
- Model output: Feedback, score and suggestions
Résumé analysis
- System: Upload a PDF and a job title
- Code: Text extracted
- Model output: ATS score, summary, improvements
Simplified diagram.
Around the interviewer sit two more tools:
- AI résumé analysis that scores a résumé for applicant tracking systems against a job title and suggests improvements.
- A mock multiple-choice exam module that draws 10 questions by difficulty and category, scores them instantly and explains wrong answers. It reduced manual evaluation by 70%.
Technology and why
- OpenAI
- A GPT-4 agent built with the OpenAI Swarm framework runs the voice interview, generating questions for the chosen role and follow-ups from the answers.
- Whisper and TTS
- Whisper turns spoken answers into text, and text-to-speech gives the interviewer a voice.
- Gemini
- Gemini 2.5 Flash reads an uploaded PDF résumé against a job title and returns an ATS score, a summary and improvements.
- FastAPI
- The API for authentication, quizzes, résumé analysis and voice sessions, with interactive docs out of the box.
- PostgreSQL
- Stores users, questions, quiz results, résumé analyses and interview sessions through SQLAlchemy 2.0.
- Redis
- Blacklists JWTs on logout, so a stateless token can still be revoked.
- React
- React 18 with Vite and Tailwind CSS v4 for the candidate-facing app.
Hard problems
Logging out of a stateless session
JWTs are valid until they expire, which is a problem for logout. PrepLens keeps revoked tokens in Redis, so a logged-out token is refused, with bcrypt-hashed passwords and 20-minute token expiry.
Results
- Less manual evaluation, thanks to instant MCQ scoring
- 70%
Model and evaluation
- Model
- GPT-4 (OpenAI Swarm agent) for the voice interviewer, Whisper for speech-to-text, text-to-speech for its voice, and Gemini 2.5 Flash for résumé analysis.