How the interviewer opens it
“Design a support system that predicts why a customer needs help from their previous Amazon orders. The system connects the customer to an appropriate support agent. The customer and the agent communicate through chat.”
The prompt is vague on purpose. The first minutes are for you. Reduce the prompt to the core features and agree on the numbers before you draw.
What to clarify first
- Customers and order context
- Support demand
- Read and write traffic
- Issue prediction
- Agent constraints
- Latency
- Consistency and availability
- Chat durability
- Privacy
- Infrastructure and scope
The interviewer has the answers. In a SysRound round, Alex gives these answers, as a real interviewer does.
Scope of a strong answer
Core requirements
- Suggest an issue and an order from the customer's recent orders.
- Let the customer confirm or correct the suggestion.
- Connect the customer to an available agent with the required skills and language.
- Support durable chat and reconnect after a connection failure.
Usually out of scope
- Build the order or payment system.
- Issue automatic refunds.
- Support voice or video calls.
- Build a general model training platform.
What interviewers ask about
- Issue prediction
- An order state does not prove customer intent. A wrong prediction can delay help or reveal data that the customer cannot access.
- Agent assignment
- Many requests compete for limited agent capacity. Disconnects and delayed responses can leave the system uncertain about who owns a conversation.
- Chat delivery and recovery
- A connection can fail between storage, acknowledgement, and delivery. Retries must preserve accepted messages without creating visible duplicates or disorder.
- Sudden demand spikes
- Servers can grow faster than the agent pool. New requests, open connections, and retries can consume resources that existing chats need.
Common mistakes
- The design includes a predictor but omits agent capacity and chat recovery.
- The system treats the predicted issue as the customer's confirmed intent.
- The system reads every past order when a small recent snapshot meets the requirement.
- The system uses cached agent presence as the final capacity check.
- The design treats WebSocket delivery as durable message storage.
- The candidate adds servers without calculating how many customers must wait for agents.
The bar by level
- Mid-level (L4/E4)
- Defines the customer flow, APIs, and core records. Supports correction, atomic agent assignment, durable chat, and reconnect. Recognizes the agent capacity limit.
- Senior (L5/E5)
- Explains reservation expiry, late acceptance, duplicate requests, and message order. Calculates spike demand. Protects existing chats and chooses databases from access patterns.
- Staff (L6+/E6+)
- Defines ownership and failure boundaries across assignment and chat. Plans zone failure capacity and fair service during overload. Connects prediction quality, privacy, and controlled releases.