About Pitchly

Inspiration

Interview preparation is broken. Most people practice alone—talking to mirrors or awkwardly role-playing with friends. We wanted to create a realistic AI interviewer that sounds human, asks dynamic questions based on your resume, and gives instant feedback.

What We Learned

• Real-time AI Streaming: Implementing low-latency voice conversations with GPT-4 and ElevenLabs
• Audio Pipeline Architecture: Building a queue system for seamless recording, transcription, and playback
• Prompt Engineering: Crafting prompts that generate natural interview questions while maintaining context
• Performance Analytics: Calculating WPM and filler word density in real-time

How We Built It

Tech Stack: Next.js + Tauri + Express + MongoDB Atlas + OpenAI + ElevenLabs

1. Built voice pipeline (Whisper STT + ElevenLabs TTS)
2. Implemented streaming GPT-4 conversations with personas
3. Migrated to MongoDB for persistent storage
4. Created performance analytics engine
5. Packaged as desktop app with Tauri

Key Challenges

Audio Queue Management

Problem: Audio clips overlapped when AI generated responses too quickly.
Solution: Built a queue system that tracks playback state and handles chunks sequentially.

AI Response Loop

Problem: AI responded to its own output, creating infinite loops.
Solution: Added strict message role validation to only respond to user messages.

Streaming Latency

Problem: Full-text TTS had 3-5 second delays.
Solution: Switched to sentence-level streaming, reducing latency to <1 second.

What's Next

a

• Solana blockchain credential verification
• Multi-language support
• Video interviews with body language analysis
• Mobile app for on-the-go practice

Built for Hacklahoma 2026 🚀

Built With

• blockchain
• elevenlabs
• express.js
• from
• git
• gpt-4o-mini
• mongoose
• next.js
• nosql
• npm
• openai
• solana
• speech-to-text)
• tauri
• text-to-speech)
• whisper
• zod