The spark
Customer support often spans a policy question, an order, a refund, and a frustrated customer. I wanted to explore an experience that could respond through voice or text while making its behavior observable to a PM.
The experience
The prototype uses fictional retail customer profiles and order data. A customer can ask about returns, refunds, or policy in English or Hindi. The interface includes customer context, past conversations, and a voice and text interaction. The PM dashboard summarizes recorded conversations, including sentiment and evaluation results.
Product decisions
- Use explicit policy and case data for answers rather than imply access to a live order system.
- Give customers a clear path to escalation when the request cannot be resolved safely.
- Make evaluation scores explainable from stored conversations instead of presenting decorative numbers.
- Show the customer context and conversation history so the experience has continuity.
How it works
The demo is a Next.js and React prototype with hardcoded customer cases and policies. It follows a speech-to-text, policy and intent handling, response, and text-to-speech pattern. The answers are rule based and use lexical retrieval over fixture data; it is a prototype, not a production customer service integration.
What I’d test next
I’d run a structured set of representative conversations across languages and intents, review incorrect policy answers and handoffs, and use the findings to improve the flow. I’d also test how clearly customers understand the agent’s limits.