Whenever you're ready, walk me through how you'd approach this.
Design a Google Reviews-like system.
Asked in the ML system design stage. The discussion covered the NLP pipeline, a BERT deep dive, LLMs versus smaller models across latency, cost, and quality, plus unsupervised clustering and summarization.
Address the end-to-end architecture, data and labels, model choices, online and batch processing, evaluation, serving, and failure modes. Explain when you would use BERT, an LLM, or a smaller model, and clarify requirements before making scale assumptions.