So how can I improve this?

As mentioned in previous posts, the program works, but it can certainly be improved. And here are some possible ways how that can be done:

• Avoid searching for the pieces in every board: When the moves are generated for a board, every square on the board is checked to see if it contains a piece that can make a move. I imagine quite a bit of time could be saved if I attached to the board a list of all the squares containing the white pieces, and all the squares containing the black pieces.
• Move ordering: A core part of alpha-beta pruning is not following up a move if it isn't going to be better than previous ones. This works much better when the first moves checked are most likely to be the better moves.
• Opening moves library: This would put the AI in a much better position for the rest of the game; using tried and tested opening moves, because it certainly cannot look forward enough itself to work out good opening moves.
• Endgame moves library: Although it performs fairly well and quickly in the endgame phase (because there are a small number of moves to analyse), it would be made as close to unbeatable as possible if some sort of endgame library was introduced when there are, say, 5 or fewer pieces. This link gives an example of what I'm talking about.