A fitness AI built from scratch — our own tokenizer, our own transformer, our own training loop. No API behind it. Ask it about training, recovery or nutrition and it answers from what it was taught; for facts it looks them up rather than guessing.
Training, recovery and food. Off-topic questions get sent back — it's a specialist, not a know-it-all.
312 million parameters, pretrained on 40 billion tokens on a desktop GPU. Every weight is ours.
Facts come from a source and are quoted, not improvised. If it can't verify, it says so.
Anything that sounds like an injury or a medical red flag gets sent to a real person, not answered.
Built by Steven McKinney at Baller Knowledge — the in-house AI behind BKoached. It runs on a single consumer GPU, so it's only awake when that machine is. It gets things wrong sometimes; that's the honest state of a model this size, and why the checking matters more than the size. Not medical advice — see a GP or physio for anything that hurts.