Building the whole path, not just calling an API
Salty Steak is a local AI workspace I am building for Windows. It brings chat, coding, model management, and training tools into one desktop app. The work behind it includes tokenization, data pipelines, training runs, checkpoints, and local inference.
I started it because I wanted to understand how the pieces fit together when there is nowhere to hide behind a hosted model. Tokenization affects the data. The data affects training. Saved checkpoints have to be valid and reproducible before inference can be trusted. The desktop app has to report what actually happened instead of guessing.
Still in progress, honestly
This is an active technical experiment, not a finished production model. Training runs produce checkpoints that can be used for inference, but I do not present the project as state-of-the-art, a ChatGPT competitor, or novel research.
That unfinished status is part of the point. The project is a record of learning the system by building it, finding the weak spots, and improving each part instead of making a five-minute demo look more complete than it is.