Then he built it in fourteen days.
The ambition was not.
What if anyone building a product could have a team of five AI specialists — each brilliant at one thing — working alongside them? Not a chatbot. Not a general-purpose assistant. Five focused minds, each trained to think the way the best product teams think about their specific craft.
One to mine research for meaning. One to frame the real problem. One to map every screen. One to critique with surgical precision. One to perfect every word.
Not tools. Colleagues.
But here's what makes keya different from another wrapper around a language model: 25 years of lived experience are baked into every agent. The way Mine structures personas isn't how Claude would do it out of the box — it's how a senior researcher who has run hundreds of studies would do it. The way Crit ranks severity isn't generic — it reflects the judgment calls a design director makes after seeing a thousand shipped products break in the same ways.
Every prompt, every output structure, every handoff protocol carries the scar tissue of real projects — the ones that shipped, the ones that failed, the ones that taught hard lessons about what matters and what's noise. Claude provides the intelligence. The intention comes from a quarter century of building products for real people.
Everything starts.
The first commit lands at 10:49 AM. By noon, there's a monorepo, a database, authentication, and an API. By evening, Crit — the design critique agent — is alive. Streaming responses. Persisting to Postgres. Tracing every thought through Langfuse.
Not a prototype. Not a mockup. A working product.
The next day brings the workspace, the dashboard, the sidebar, project navigation, Stripe billing, and credit-based pricing. The app shell takes shape — dark, minimal, every pixel considered.
Day three: all five agents ship. Mine synthesizes research into personas and insight clusters. Frame writes design briefs with territories and target journeys. Plot maps screen flows with Figma-ready prompts. Tone audits and rewrites every string.
Three days. Five agents. Zero shortcuts.
They hand off.
The breakthrough wasn't making five agents. It was making them work together. Mine's personas flow into Frame's brief. Frame's territories feed Plot's screen maps. Crit's findings shape Tone's rewrites.
Each handoff carries full context. When Frame receives Mine's output, it doesn't just get text — it gets structured personas, ranked insights, and opportunity statements. The receiving agent understands what came before.
And the flow isn't rigid. Crit can send you back to Plot. Tone can flag problems for Crit. The agents suggest, they don't dictate. Because real design work isn't linear — and keya knows that.
that nobody asked for.
A four-strategy JSON parser that handles fenced blocks, bare JSON, full-output JSON, and truncated output repair. Because when a model hits max tokens mid-sentence, you don't lose the user's data. You close the brackets and keep going.
Every user-facing string in one file. Change the voice of the entire app by editing copy.ts. Not because someone asked for it. Because the product might need to speak differently someday, and the architecture should make that trivial.
Project settings that become agent personality. When you tell keya "this is a healthcare app for elderly users," the agents don't just filter — they change how they think. What they emphasize. What they flag. Settings aren't configuration. They're character.
A pre-run confirmation that turns a black box into a collaborator. "Based on your input, I'll produce 3–4 behavioral personas and insight clusters. This will use 2 credits. Proceed?" Because a good colleague confirms scope before starting work.
the hard look.
On day 11, a UX audit identified 28 gaps. The app was feature-rich but designed around what the system could do, not what the user was trying to accomplish. Seven clicks before first value. Mode toggles that exposed system complexity. Five agents shown to users who could only use one.
The honest assessment: "The highest-impact fix transforms the agent from a black box into a collaborator."
So the toolbar disappeared. Mode selection became automatic. The project history moved inline. Post-run nudges started guiding users to the natural next step. The suggested agent's tab began to glow.
The app didn't get more features. It got more thoughtful.
actually produced.
Five AI agents with structured output, streaming responses, project-level persistence, and cross-agent handoffs.
A complete platform — auth, billing, credits, workspaces, projects, settings, templates, export, sharing, history.
Security and compliance — per-user rate limiting, GDPR cascade delete, session management, CORS lockdown, Terms of Service, Privacy Policy.
A guided experience — pre-run confirmation, post-run nudges, tab glow, credit toasts, contextual handoff suggestions.
And 45 bug fixes. Because shipping isn't building features. It's making them work for real people with real expectations.
at IDYeah Studio
with Claude Code
Named after his daughter.
Built with love.