Home > Projects > soulsaka – Personal AI Clone

soulsaka – Personal AI Clone

2026 · Working; retrains monthly on my own machines

A personal AI clone that runs entirely on my own computers. It imports my message history, fine-tunes a small language model on how I write, clones my voice from my own recordings, and keeps score of how close each version gets to the real me.

Tools: Python · TypeScript · PyTorch · Local LLMs · Voice cloning

GitHub

More info

What it does

Imports my history with one command: iMessage, WhatsApp, email, Discord, git commits, my own writing. Texts, emails and speech are tagged separately so the model learns that I write differently in each.
Trains a small language model on everything I have written, and retrains from scratch every month so versions can be compared.
Clones my voice from a short reference clip assembled from my own recordings.
One machine can keep its microphone on all the time; it keeps my voice and discards everyone else's speech.
Memories: say "remember the locker code is 4521" on any device and it is on every device within seconds. It also pulls useful facts out of normal conversation.
Measures every version: friends guess which of two replies is the real me, a classifier tries to tell my text from the model's, and the cloned voice is compared to mine.

Architecture

A hub (the computer with the GPU) serves the model and the API. Phones, laptops and browsers pair with the hub by code and sync offline-first. Chat uses the local model by default; a cloud model can be plugged in with an API key if I choose to.

Nothing leaves my machines. That is the whole point.

« Back to all projects