Build Your Own AI From Scratch — Trained on Your Own Texts

"Can I create my own AI like ChatGPT?" is one of those questions everyone asks and almost nobody actually tries, because the path looks like a wall of jargon. This repo — texts-to-transformer — is the closest thing to a friendly on-ramp I've found.

It trains a tiny transformer from scratch on one very personal dataset: your own text messages. No fine-tuning someone else's model, no cloud, no API keys. You build the thing from zero and end up chatting with a little model that writes like you.

1What it does, step by step

The project is a full pipeline — nine steps — and it's honest about every one of them:

2Can you really make your own AI without coding?

Not quite — this one asks you to run a Python pipeline, so you'll touch a terminal. But you don't need to write the model; it's all there. If you can follow steps and paste commands, you can get through it. And honestly, doing it this way teaches you more in an afternoon than a month of watching videos about it.

3Is it private?

Completely. That's the whole point. Nothing leaves your laptop — not the messages, not the model. For a dataset as personal as your own conversations, that matters a lot, and it's why training locally with MLX is the right call here rather than uploading everything to some service.

4What you actually get out of it

Two things. First, a strange little mirror — a model that has picked up your phrasing, your habits, the way you actually type. It's uncanny in a fun way. Second, and more useful: you'll finally understand, from the inside, how a transformer turns text into more text. that understanding transfers to every big model you'll ever use.

5How to actually build it

This is the part the title promises, so here's the real path. You need an Apple-Silicon Mac (M1 or newer), macOS 14+, about 16 GB of RAM, and Full Disk Access granted to your terminal.

git clone https://github.com/Doriandarko/texts-to-transformer.git
cd texts-to-transformer
brew install uv        # skip if you already have it
uv sync
uv run imessage-mlx doctor

The doctor step checks Apple Silicon, MLX, disk space and — importantly — that it can read your Messages database safely before anything touches real data.

uv run imessage-mlx snapshot --config configs/data.yaml
uv run imessage-mlx prepare  --config configs/data.yaml
uv run imessage-mlx privacy-audit

snapshot copies your chat.db in read-only mode (the original is never touched), and the pipeline pseudonymizes contacts before writing anything. Then you train and talk to it:

uv run imessage-mlx train-tokenizer --train work/splits/train.jsonl
uv run imessage-mlx train
uv run imessage-mlx export
uv run imessage-mlx chat --model outputs/final

Exact flags for each step live in the repo's README — I kept the commands here to the core sequence so you can see the whole shape at a glance.

324 stars in a week says a lot of people have been quietly curious about a personal AI that knows how they write. If that's you, and you've got a Mac, this is a genuinely great weekend project — build your own AI, from scratch, on data only you have.