PrepTalk

Inspiration

Job interviews can be nerve wracking, especially for students and early career professionals. We wanted to create a tool that makes interview practice accessible, realistic, and personalized. By leveraging AI technology, we aimed to provide instant, constructive feedback that adapts to each user's unique resume and target position, something traditional mock interview resources can't easily offer.

What it does

PrepTalk is an AI powered desktop application that conducts realistic mock job interviews. Users upload their resume (PDF) and paste a job description, then receive personalized interview questions generated by Gemini AI. The app records spoken responses via microphone, transcribes them using OpenAI's Whisper model, and provides real time AI feedback on answers. After completing multiple questions, users receive comprehensive feedback on their overall performance, including strengths, areas for improvement, and specific examples of good and bad responses.

How we built it

We built PrepTalk using:

• Frontend: PyQt6 for a modern, gradient styled desktop GUI with smooth page transitions
• Backend: Python with modular architecture separating audio processing, AI conversation management, and system prompts
• AI Models: Google's Gemini 2.5 Flash for question generation and feedback, OpenAI's Whisper (medium) for speech to text transcription
• Audio Processing: sounddevice and soundfile for microphone recording, NumPy for audio array manipulation
• Document Parsing: pdfplumber for extracting text from resume PDFs
• Configuration: YAML based system prompts for easy AI behavior customization, python dotenv for secure API key management

Challenges we ran into

• Audio Recording Reliability: Getting consistent microphone recording.
• AI Prompt Engineering: Crafting system prompts that generated realistic, relevant interview questions without inventing candidate experience not in the resume required iteration.
• Whisper Model Performance: Balancing transcription accuracy with processing time. We settled on the "medium" model as the sweet spot.
• Cross platform Audio: Handling different microphone configurations and permissions across Windows, Mac, and Linux.

Accomplishments that we're proud of

• Successfully integrated three major AI technologies (Gemini, Whisper) into a cohesive user experience
• Created a clean, modern GUI that feels professional and intuitive
• Developed a flexible system prompt architecture using YAML that makes the AI interviewer's behavior highly customizable
• Built a fully functional MVP in a hackathon timeframe with proper documentation and licensing

What we learned

• How to effectively use AI models for complex workflows (document parsing to question generation to speech recognition to feedback)
• The importance of structured system prompts in controlling AI behavior and preventing hallucinations
• PyQt6's capabilities for creating modern, styled desktop applications
• The value of modular architecture.

What's next for PrepTalk

• Enhanced Frontend: Improve UI/UX with animations, better visual feedback, and a more polished design
• Text to Speech Functionality: Add voice output for the AI interviewer to read questions aloud, making the experience more immersive and realistic
• Customizable Question Count: Allow users to select the number of interview questions they want to practice with
• Multi turn Conversations: Enable follow up questions based on candidate responses for more realistic interviews
• Video Recording: Add webcam support to analyze body language and presentation skills
• Industry Specific Modes: Specialized question banks and feedback criteria for technical, behavioral, or case interviews
• Web Version: Port to a web application for broader accessibility without installation
• Integration with Job Platforms: Automatically fetch job descriptions from LinkedIn, Indeed, etc.

Built With

• python