Home

IRI

Children's Korean voice AI with QLoRA-tuned Kanana 3B on Runpod GPUs, served through FastAPI

IRI voice interface with a pastel orb and microphone and playback controls

IRI lets children aged 4 to 10 ask questions by voice or text and hear Korean answers adapted to their age.

Fine-tuned Kanana 2 3B Instruct with QLoRA in PyTorch for conversations with children. Built a Python backend with FastAPI to connect speech transcription, model responses, and speech synthesis. Deployed Kanana on Runpod GPUs and added a fallback model for requests when the GPU server is offline.

Designed and built the voice interface and landing page in React and TypeScript with Vite. Animated a Three.js orb in response to voice activity and the listening, response generation, and speech states.

React TypeScript Vite Three.js Python PyTorch FastAPI