TEXTE

I Built a Voice AI Concierge for My Portfolio

10. August 2026 · 3 Min. Lesezeit

This site has a voice you can talk to. It runs on two pipelines — VAPI for managed voice, NVIDIA PersonaPlex on a rented GPU for real-time full-duplex — and hands off between them mid-conversation. Here's why I built it and how the pieces fit.

What It Takes to Run a GPU Voice Model on a Serverless Budget

4. August 2026 · 3 Min. Lesezeit

PersonaPlex needs a GPU; my site runs on Vercel serverless. Bridging the two meant a RunPod pod lifecycle with a distributed lock, heartbeats in Redis, a cron-driven reaper, and a watchdog inside the container that kills the pod when visitors leave. Here's the honest version, tradeoffs included.

Teaching an AI to Answer Questions About My Career

28. Juli 2026 · 3 Min. Lesezeit

The concierge on this site answers from a 2,600-line brain dump of my career, chunked and searched in memory — with pgvector as the planned upgrade. This is how the knowledge layer works, how the prompt keeps it honest, and where it still fails.