anapaolaoviedo

Ana Paola
Oviedo

Mathematician · Theorist · Builder

I study the foundations of computation and logic — then build the things most people just import.

JARVIS voice loop — on-device pipeline schematic input · 16 kHz WAKE-WORD CLASSIFIER raw MFCC features · training ASR mlx-whisper · on-device REASONING local LLM · ollama state TTS voice out on-device · zero cloud calls
fig. 01 — the JARVIS loop, as built
now building a lab notebook, kept in public

JARVIS — local-first AI assistant

in progress

A voice assistant that runs entirely on-device — no cloud APIs, no black boxes. Every model in the loop runs on my own machine, and every stage I can't explain, I rebuild until I can.

  • 2026.08Voice loop working end-to-end: on-device ASR (mlx-whisper) → local LLM (Ollama) → TTS
  • thenIntent parsing & planning
  • thenAction execution
  • thenComputer vision

Named after the one in Iron Man, obviously. Tony Stark built his in a cave; I'm building mine on a MacBook.

full project notes → why train my own wake-word model →
"Computer science is not about coding. It never has been. It is about human thought. Can machines be kind? That is my question."
— the question underneath everything here