Built in one evening, on my own machine

I built an AI avatar
of myself.
It never left my house.

I wanted to know how much of this could actually be done privately. Not through a subscription or an account somewhere else, but here, on a computer I control.

So I tried it.

A face, a voice, and a body of knowledge, assembled in my spare room. Here is what I built, what it can do, and where it still falls short.

This is an AI avatar. The face and voice are mine and the words are mine, but I did not sit down and say them. It says so itself, at the start, because I think anything made this way should.

What you just watched

What you just watched is actually three different things, all running on the same machine and brought together at the end.

The voice

Cloned from seventy-four seconds of me talking into a microphone. It can now say anything, in about two seconds a line.

The face

My own video footage. The model changes only the mouth, so everything else — posture, light, the room — is really me.

The knowledge

A twin that has read my own work and answers from it, quoting real prices and citing which page it took them from.

The first two make it look like me. The third is what allows it to know something about me.

And for me, that is the more interesting part.

What it cost

Voice cloning$0 — open model, runs locally
Face and lip sync$0 — open model, runs locally
The language model$0 — open weights, runs locally
Ongoing subscriptionnone
Data sent to anyonenone
Hardwarea desktop computer I already owned

There are commercial services that can do versions of all of this, and many of them are very good. What interested me was finding out how much could happen without my material ever needing to leave a machine I control.

Why this is really about privacy

The avatar is the visible part. But it isn’t actually the reason I built this.

I wanted to know whether capable, genuinely useful AI could live close to the person it belongs to. Whether it could work with someone’s voice, face, writing, documents and knowledge without requiring all of that material to be handed over somewhere else.

This experiment gave me an answer.

It can.

And if your work holds other people’s stories, private information, or confidences, that begins to matter very quickly. What a private setup looks like for practitioners is the longer version.

What it still can’t do

This isn’t magic, and it isn’t finished technology.

The lips are close, but not exact. If you look carefully, you can tell the video is generated. Longer videos take hours rather than minutes. And the knowledge system only knows what I have actually given it. When it doesn’t know something, I want it to say so. That last part took some work.

I don’t want the demonstration to be better than the truth.

If I am going to show you what this technology can do, I also want you to know where its edges are.

If this makes you curious

I build systems like this for other people too, on hardware they control, using their own material.

If you are wondering what private AI could look like for your work or practice, we can simply talk about it. The first conversation is free, and if I don’t think this would actually be useful to you, I will tell you that too.

Book a free twenty minutes