Tools
Workflow
End to End
Knowledge base
AI
Git
API
Project overview
An AI voice agent built for this portfolio, so visitors can just talk to it instead of reading through everything. It listens, thinks, and replies out loud in real time, grounded in a knowledge base about my work so it stays on topic and doesn't make things up. Built from scratch with prior Git, Node.js, API.
Why?
Most of this portfolio is text and images to scroll through. I wanted a section people could just talk to, and a chance to learn a full AI voice pipeline hands-on, speech in, reasoning, speech out, with prior backend experience.
5-stage pipeline
Deepgram, GPT-4o-mini, Supabase pgvector, ElevenLabs, and Vercel woven into one real-time voice loop.
Built from zero
With Git, Node.js, or API experience going in, the whole stack learned and shipped end to end.
How it works
A Siri-style orb that listens.
Speech in, speech out, in real time.
Grounded in a real knowledge base.
What I Learned
Shipping this meant learning Git, Node.js, and API work from a standing start, and debugging real production issues along the way: a malformed Supabase URL, ElevenLabs 402 errors from premium voices, hardcoded name references in the system prompt, and a Safari/iOS audio playback bug fixed with a pre-unlocked persistent audio element. I also tried a more minimal redesign of the widget, a bare orb with an ambient edge glow, and ended up rolling it back in favour of the original.
