Inside our AI boot camp: what thirty participants built in five days

Five days, thirty participants, and one rule: every team ships something that works on real hardware by Friday afternoon. This year's AI boot camp produced nine working prototypes, and more importantly, nine honest stories about what broke on the way.

The standout build was a crop-disease detector trained on photographs taken by the participants themselves in the fields around campus — small dataset, careful augmentation, and a model small enough to run on a phone. Close behind was an attendance system that recognises faces entirely on-device, with no image ever leaving the room.

Not everything went smoothly. Two teams lost a day to a mislabeled dataset, one team discovered on Thursday that their model performed beautifully on their own faces and poorly on everyone else's, and the robotics-adjacent team learned that inference latency matters more than accuracy when a motor is waiting on your answer.

That is exactly why we run the camp this way. The demo matters, but the debugging matters more — and every participant left with a project they can explain line by line, because they had to fix it line by line.