1
1
Looking for experience!
.png&w=828&q=75)
Bob The Builder uses IBM Bob as an AI research engineer to auto-optimize CIFAR-100 image classifiers. Bob runs 30 systematic experiments testing learning rates, batch sizes, optimizers, dropout, batch normalization, weight decay, and data augmentation. Each iteration, Bob changes one hyperparameter, trains the model for 10 epochs, reads the accuracy, compares with best result, commits improvements with Git, and rolls back failures with Git reset. A Streamlit dashboard shows real-time progress. Baseline accuracy was 40.43%. Bob achieved 44.67% - a 4.24% improvement. Batch Normalization was the winning change. Dropout, weight decay, and SGD optimizer all failed. Bob proved that simplicity wins for CIFAR-100. All code is on GitHub. All 30 experiments are logged.
17 May 2026