WuKong AI

Inspiration

AI research papers are incredibly dense. Even experienced engineers struggle to understand breakthrough papers like DeepSeek's Multi-Head Latent Attention. I wanted to create a way to learn complex AI concepts through natural conversation - just like having a knowledgeable tutor available 24/7.

What it does

WuKong AI transforms complex AI research papers into interactive voice conversations. Users can:

• Ask questions naturally - "What is Multi-Head Latent Attention?"
• Get instant explanations - AI tutor responds in real-time voice
• Learn at their own pace - Three-layer content: summary, visuals, and code

How I built it

• LiveKit Agents SDK for the voice AI pipeline
• LiveKit Cloud for deployment and real-time WebRTC streaming
• LiveKit Inference for STT (Deepgram), LLM (GPT-4.1), and TTS (Cartesia)
• React + TypeScript for the frontend
• Python for the agent backend

Challenges I ran into

• Configuring the voice pipeline to feel natural and responsive
• Designing prompts that make the AI explain complex concepts clearly
• Building a complete product solo in 48 hours

Accomplishments that I'm proud of

• Built a fully functional voice AI tutor from scratch
• Deployed on LiveKit Cloud with 24/7 availability
• Created an engaging learning experience for DeepSeek MLA

What I learned

• LiveKit Agents SDK is incredibly powerful for building voice AI
• The importance of turn detection and natural conversation flow
• How to structure AI explanations for complex technical content

What's next for WuKong AI

• Add more AI paper case studies
• Integrate Overshoot for visual recognition (scan paper screenshots)
• Build a content library for the AI research community

Built With

• livekit
• lucide
• next.js
• react
• shadcn-ui
• stripe
• tailwind-css
• typescript