End of Volume Review & Interview Questions

Three proposals arrive for the same problem. An insurer wants to route each incoming claim description to one of 30 handling teams. The first team proposes a two-layer LSTM trained from scratch, because claims are sequences of words. The second proposes fine-tuning a pretrained encoder, because routing is classification and BERT-style encoders were built for it. The third proposes fine-tuning Llama 3.1 8B, because it is the most capable model the company can run itself. Each proposal cites a chapter of this volume, and each is defensible in a sentence.