Overview

Icons is a premium B2C AI voice coaching platform (iconsclub.com). The build had been stalled for ~1.5 years across legacy and vendor design when I joined. I audited 1,000+ screens block-by-block against the engineering roadmap, mapped every design artifact to actual API capabilities, and replaced static mock-based UI with a data-driven screen architecture. I designed the full real-time voice AI coaching UX and shipped the subscription paywall to production.

The Ask

Un-stall an 18-month-old design legacy, reconnect it to what the product can actually do, and get a subscription business live before first release.

Constraints

• ~1.5 years of stalled legacy & vendor design

• Static mocks disconnected from actual API capabilities

• Real-time voice AI: latency, streaming, interruption states

• Premium brand bar with a pre-GA launch deadline

• App Store launch mechanics — launch-boost risk flagged to PM

What I Did — UX Architecture

The defining reframe: the interface is a real-time conversation surface, not a set of screens. Every state is derived from what the system can actually do at that moment — streaming, partial results, degraded connectivity, interruption. The screen architecture is data-driven, mapped 1:1 to API capabilities, so design and engineering share one source of truth.

Interaction Design

Interaction design focused on trust in a live AI conversation: clear listening / thinking / speaking states, graceful interruption handling, and coaching context always visible — the user knows what the AI knows. The subscription paywall shipped to production in RevenueCat Paywalls v2: pixel-matched to Figma, natively rendered, remotely configurable, and A/B-test-ready.

Key Screens

• Real-time voice session: streaming, interruption, and recovery states

• Coaching context display: what the AI knows and why it responds

• Subscription paywall: RevenueCat Paywalls v2, live in production

• Marketing website: WCAG 2.1 AA

Outcome

Subscription paywall live in production; WCAG 2.1 AA marketing site shipped. The product is now moving to its first public release — recovered from an 18-month stall.

What I'd Do Differently

Push for the API-capability mapping exercise in week one rather than week three — half the legacy mocks turned out to reference endpoints that no longer existed.