Skip to content
Jaswanth Samba
All projects

01 / 04Private voice assistant

Luna

A voice assistant that runs entirely on my laptop, with no cloud calls.

  1. 1Wake word"Luna"
  2. 2Speaker recognition
  3. 3Llama 3.2 3Blocal model
  4. 4Kokoro speech
Everything on this path runs locally.

The problem

To doWrite the real story behind Luna in one short paragraph: what made you want a voice assistant that never leaves your laptop?

Luna is privacy-first by design. Every step, from hearing the wake word to speaking the answer, runs on my own machine with no cloud calls.

What it does

How it works

  1. 1Wake word"Luna"
  2. 2Speaker recognition
  3. 3Llama 3.2 3Blocal model
  4. 4Kokoro speech

Alongside the voice loop, Airflow pipelines keep financial data up to date, and every utterance is embedded into a semantic memory that the model can search.

To doAdd the full architecture diagram (voice loop, memory store, Airflow pipelines, dashboard) and a short explanation.

What was hard and how I solved it

To doDescribe the hardest problem (for example, keeping response time low on a laptop GPU) and how you solved it.

Results

llama3.2:3b on an RTX 3050 Ti laptop GPU
59 tokens/s
llama3.2:3b on an RTX 3050 Ti laptop GPU
Airflow pipeline runs succeeded across 3 pipelines
7 of 7
Airflow pipeline runs succeeded across 3 pipelines
of logged utterances embedded for semantic memory
100%
of logged utterances embedded for semantic memory
To doLog response latency for a day and add the median "hear you to first sound" time.
To doRecord a 30 to 60 second demo video and add a dashboard screenshot. Never publish household audio or conversation logs.

Stack

Links

To doGitHub link. Repo is clean (only .env.example tracked) but has no remote yet.To doDemo video