The hard part isn't just "can an LLM take an order" — it's making the whole conversational loop (does the customer sound done talking, was the transcript actually right, did the model behave consistently across a thousand near-identical calls) hold up against real noise, real accents, real mumbled orders, and a live customer who won't repeat themselves. As an AI Researcher, you'll push on the modeling and methodology underneath that loop: turn taking, voice activity detection, speech accuracy, order accuracy, detecting side conversations, and the rigorous experimentation needed to know — with statistical confidence, not vibes — whether a change actually made things better.