Capstone: MegaMart Assist, an AI support assistant, end to end

The course's last interview: design an AI customer-support assistant that answers with citations, acts on orders and hands off to people, in English, Hindi and Hinglish. Requirements and a token estimate (23,010 input tokens a conversation, ~3.45 million counted tokens a minute uncached, ~290,000 cached); cost per conversation; a controller, a tool gate and an LLM gateway; one turn through the agent loop; a first-words budget of about 0.9 to 1.3 s; hybrid retrieval in pgvector with filters inside the search; fresh, versioned, permission-aware ingestion; narrow tools where the model never picks the refund amount; proposals the customer confirms, executed once; injection defences; evals and rollout; a degradation ladder that keeps the tools identical; the prompt layout, routing and budgets; the Hindi sale-evening curveball; SLOs and dashboards; and a send-off for the whole course.