Design an AI Voice Chat Assistant for a Sneaker Retailer's Website

Design a voice chat system for a retailer's website. The example given was the online store of a sneaker and sportswear chain such as Foot Locker: a shopper on the site talks to an AI assistant, and the assistant answers out loud. The prompt is deliberately open and tests product decomposition: decide what the product should do, break it into components, and justify the trade-offs yourself. Asking the interviewer how something should work gets the question turned back to you, so propose an answer and move on. ```hint Pick the first conversations List the few shopper requests that would make talking to the store worth it, and decide which of them need the shopper to be signed in. ``` ```hint Spend the latency budget Write down every step between the shopper finishing a sentence and hearing the first word of the reply, and decide which steps can run at the same time. ``` ```hint Where the facts come from Consider what the assistant may say about price, stock, sizes and orders, and which system must be the source of each of those statements. ``` ### Clarifying Questions - Is the assistant only a voice widget on the website and app, or must it also answer the customer-service phone line? - Which journeys come first: finding products and sizing advice, checking stock by size or store, order status and returns, or buying by voice? - Which existing systems can it call: catalog search, inventory, order management, store locator, the support desk? - Which languages and markets must the first version support? - Which actions may the assistant take on its own, and which need an on-screen confirmation or a human agent? - May conversations be recorded and transcripts kept, and under what consent and retention rules? ### What a Strong Answer Covers - Ownership of scope: named shopper journeys, a prioritized first version and explicit non-goals, chosen without waiting for the interviewer - A component breakdown of the voice loop: audio capture and playback, streaming speech recognition, end-of-turn detection and interruptions, an LLM agent with store tools, speech synthesis and session state - A latency budget for natural conversation and the streaming and overlap that achieve it - Grounding and safety: prices, stock and order facts only from tools, sign-in before account data, confirmation before actions, resistance to off-topic requests and injected instructions - Human handoff with context, and privacy handling for audio and transcripts - How success is measured before and after launch, and the cost and capacity plan for traffic peaks ### Follow-up Questions - The shopper interrupts the assistant mid-sentence to change the size they asked about. What happens in each component, and what goes into the conversation history? - A limited sneaker release sends a sudden surge of shoppers to the site. What do you scale ahead of time, and what do you degrade if capacity runs out? - How would you test a new prompt or model version before it talks to real shoppers? - When should the assistant hand a conversation to a human agent, and what should that agent receive?