Between understanding an idea and shipping it

A sentiment classifier for Arabic

A customer-support team receives thousands of Arabic messages a day across several dialects. They want each one sorted into positive or negative before a human sees it.

Defect detection on the line

A production line photographs every part. The classifier must run on the line's own CPU between parts, so it gets a hard compute budget — and the plant may swap in a higher-resolution camera next year without retraining you a new head.

The stack that will not learn

A colleague hands you this model. It compiles, the shapes all line up, and it trains for hours without the loss moving. Nothing about the diagram is wrong — and that is the point.

Loss sits at its initial value. Gradients measured at the first layer are ~1e-9; at the last layer they are normal.

A residual block that survives depth

Your plain convolutional stack got worse when you made it deeper — not overfit, just worse on the training set too. The 2015 answer was to change the shape of the block itself.

The generative decoder block

Every modern language model is one block repeated dozens of times. Build that block once, correctly, and you understand the shape of all of them.

Adapting a large model to clinical notes

A hospital wants a triage model over its own clinical notes. The data cannot leave the building, and the building has one GPU. Fine-tuning every weight of a large pretrained model is not on the table.