This website is devoted to improving players' draft skill.
It uses deep learning to approximate players' decisions made during MTG draft.
AI calculates score for each card in pack based on three coefficients:
Finally, scores summed for each card in pack. Card with the highest score is chosen by AI.
This model doesn't considers some aspects of drafting like sending signals, forcing an archetype or expecting for a card to make circle and come back. It doesn't plans how to draft, it only learns to make single picks. It doesn't solves the format, instead it tries to model human player, so you can treat it as an averaged human.
For each new format model starts with some pre-initialized weights (which incentivize it to pick rares over uncommons, cards with same color). Then, with every new batch of picks from human players, it starts to refine those weights. Experience shows that model needs about 1000 drafts to converge, so don't expect it to make sane picks in some rare formats. But the more popular format is - the smarter model will be.