Designed the AI coach that lets sales reps lose the deal before it matters.

Sales reps lose deals in moments they never practiced for. There was concept on a napkin: AI-powered role-play that lets reps rehearse difficult conversations before they happen in real life. There was no product, no design direction, no existing patterns to reference. Just a concept and a deadline. I owned every product and design decision from zero without any handoff, any brief, any existing system to inherit. I defined what the core experience needed to feel like: low enough friction that a rep would actually use it before a real call, structured enough that the feedback meant something afterward. That meant designing three distinct surfaces: a scenario builder for managers to configure AI personas and objections in seconds, a live roleplay interface that didn't get in the way of the conversation, and a post-session feedback model that surfaced specific mistakes and specific strengths — not generic scores. In early beta, the product logged 39+ calls across 12 accounts without a single prompt from the team. The hardest design problem in AI products is trust. A rep has to believe the AI is giving them real feedback, not flattery, or they'll use it once and never return. Every design decision in Arena was in service of that: making the AI feel like a tough but fair coach, not a chatbot ticking boxes. The beta results suggested it was working. The next question was scale.