margin

• 

  The UI

Inspiration

AI writing tends to sound pretty awful (Claude has tell-tale signs of slop—if you know what I mean). We thought it would be pretty useful to have something that wrote just like us! It would allow for a lot of cool use cases :).

What it does

It is a plugin to your Gmail (or Slack or messaging apps or any other app you'd like) that trains on past threads (completely locally on your computer) and learns your voice. It starts from a base model so it doesn't even have to unlearn the assistant persona! And it creates multiple "personas" (professional, social, and more to be created soon!) so you can choose when to draft with a professional voice or a social one.

How we built it

The fine-tuning is done using MLX which uses kernels for Metal (this is only on Mac for now) to speed up training and use less memory. The UI uses electron to cleanly display all the generated personas.

Challenges we ran into

Training on a MacBook is pretty slow, especially for testing our app during the hackathon development, and generally doing all of this in <1 day was pretty difficult for us.

Accomplishments that we're proud of

We got a somewhat working product!

What we learned

We learned a lot about UI design, handling infra robustly, and model training.

What's next for margin

We plan to make it work on other platforms, potentially use smaller quantizations to allow for larger models, and involve other apps.

Built With

• mlx
• python
• transformers
• typescript