How Prelude's AI-native design team ships production-grade PRs, from idea to merged in 13 days

Prelude is the trust layer between signups and the business. It brings verification, auth, and fraud prevention into one onboarding decision.

A Series A company backed by 20VC, Singular, and Seedcamp, it counts BeReal, Suno, and Voodoo among its customers. The product is data-dense and dashboard-driven, built by an AI-native design team that ships its own PRs.

Like most product teams, Prelude's designers and engineers worked in separate worlds. Design explored in Figma while engineering built in the codebase, and the space between them ran on handoff: specs, revisions, and waiting. With a small design team relative to engineering, that gap had a real cost. Ideas queued behind engineering capacity, and idea to production stretched from weeks into months.

Prototyping was the first problem Prelude set out to solve. Exploration lived on mock data: frames that looked right but couldn't behave right, showing the happy path instead of the messy real states a trust-and-fraud dashboard has to handle. The team often saw how a screen really held up only after engineering had built it, when feedback costs the most. The AI design tools they tried didn't close the gap. Their outputs ignored the design system, hardcoded values instead of using design tokens, and couldn't prototype real interactions on live data.

Prelude chose Modeinspect for the thing every tool they'd tried had missed: fidelity, meaning real adherence to their design system and tokens rather than an approximation. In the pilot, ~95% of generated UI composed from their existing components and tokens, with zero hardcoded values in merged output. That's what could actually close the gap: designers exploring, prototyping, and fixing things in code directly, in the same world engineering works in, without a handoff for every iteration.

They were skeptical, and past experience had earned that skepticism. But the question was worth answering: could they change code themselves and trust the result?

The real skeptics were the engineers. Guillaume's team had watched AI tools spray slop into codebases: code that runs but reads nothing like the repo, and becomes someone's cleanup job later. They expected more of the same.

A short pilot changed their minds. The very first PR Modeinspect produced went through standard engineering review and merged. The ones that followed held the same bar, reusing their logic, following their conventions, and fitting their architecture. They were starting points engineering could genuinely build on, not the cleanup job they'd braced for.