Spec27: an engineer-built AI-validation platform, made clear.
Safe Intelligence built Spec27 to validate AI agents without teams building their own test infrastructure. The engineering was real; the interface fought its users. At North, I led the redesign across research, strategy, branding, every screen, and the dev-ready build engineering shipped from.
Live in production. The relaunch drew 500+ sign-ups in its first cohort.
The problem, and what we found
Before designing a single screen, I mapped the territory: how analogous consoles organise themselves, what the product actually was underneath, and what its users kept tripping on. The client's own material (repo threads, Looms) was mined with AI for speed and crossed with my audit. Everything landed in one feedback document that aligned client and team on what was broken.
Five findings
The structure fought how the product actually thinks: projects, agents, specs and evals were scattered rather than related.
Inconsistent, dated and poorly accessible, with components that behaved differently on every page.
Light and dark modes disagreed; the logo had no application for either. Nothing to build screens on. (It even had a different name: DeepScan.)
New users met blank tables: the product at its emptiest exactly when it most needed to explain itself.
Deeply technical language, no onboarding, no visible path to a first result.
Two deadlines, two phases
Two hard dates shaped the whole engagement: a client event weeks away, and a relaunch after it. So the work split in two: a surgical quick win first, the real redesign second.
Phase 1 · the quick win
Three touchpoints restructured without touching the brand: the home, each project's overview, and a first pass of navigation and topbar. Enough clarity, fast enough, for the client to present at the event and run user tests on something that held together. The outcome was qualitative, and it bought trust for phase 2.
Drag to compare the same product weeks apart. Phase 1 stayed inside the old brand on purpose.
The redesign
Phase 2 was the full rebuild, run on an agentic workflow with humans at every gate. Plans became design.md and roadmap.md; prototype and critique agents produced and challenged the work; the client and their engineers reviewed each step. Critique lived in chat, implementation and audit in a code agent, exploration in a design agent.
A brand that could work
Two weeks of light branding before the screens: a style guide, a motion language, and the product shell wearing the new identity, with light and dark finally agreeing and a logo that works in both.
The corrected model
The navigation only made sense once the model did. A project holds three first-class objects: agents, specifications, and the evals that pair them. Datasets and judges became plumbing: configured when needed, never navigated.
The pivot: from wizard to loop
Halfway in, spec creation changed nature. The client's own prototyping showed users don't configure once and move on; they iterate between an agent and its spec until it behaves. The sequential wizard I was designing died that week. In its place: a micro-wizard for the essentials, then a guided loop inside the spec editor where the user, not the form, decides when the spec is good enough.
Designed by states
With the 0→1 defined, every core surface was designed three times: empty, first project, and full of data. State 0 walks you straight to your first specification, with a built-in test agent always available, so integration friction never blocks the way.
State 0, nothing yet. The overview becomes an onboarding: create your first specification.
Results: one atom, three lenses
Results used to live scattered across pages that mirrored the resource pages: two “Evals” pages, robustness split by type. The redesign named the valuable datum, the spec run, and made every view a lens over it: group by eval, by agent or by spec; rows expand in place; everything crosslinks.
The consolidated Results page: one place, grouped three ways, every row crosslinked.
Delivered as a dev-ready kit: HTML, CSS and JS. Engineering shipped the rebuild in two weeks, in time for the relaunch, which the conventional route wouldn't have allowed.
Outcome
When the model is right, the product explains itself. The client wrapped the cycle happy and booked the next one.
First real cohort: sign-ups, not yet activation. The next cycle starts in September.
Don't take my word for it. Open it.
The full redesign runs in your browser: sign in, switch between the empty, first and full states, and click through the spec loop and results.
Open the live prototype ↗
