1n4NO
All work

Product Studio

UX design with checks, not guesses.

Next.jsTypeScriptCheerio (HTML parsing)Anthropic / OpenAI / Gemini APIsOllama (local fallback)

Project brief

The product bet, in plain language.

Built a product studio that moves from brief to blueprint, through deterministic UX checks, and out to exportable assets.

Hiring signal

Senior ownership across product framing, scope, and shipped interface quality.

Engineering read

Focus on constraints, architecture choices, external systems, and data flow.

Role
Product architecture, AI workflow design, design-system modeling, and UX audit implementation
Scope
Provider setup, brief capture, blueprint generation, theme tokens, audit scoring, and export artifacts
Core constraint
Generated output needed measurable quality checks instead of relying on model confidence or visual plausibility.
Technical bet
Pairing generative agents with deterministic HTML checks would make page output easier to ship.
Drag to scrub
Ollama model selection flow
Product Studio AI provider setup modal in dark mode
Provider setup modal
Product Studio workspace after Ollama setup in dark mode
Studio workspace — dark mode
Product Studio final export page in dark mode showing gate checks and export artifacts
Final export — dark mode
Product Studio AI provider setup modal in dark mode
Provider setup modal
Product Studio root landing page in dark mode
Ollama model selection flow
Product Studio audit results page in dark mode
Audit results — dark mode
Product Studio root landing page in dark mode

Overview

A browser-based UX studio: turn a brief into a structured page blueprint, check it against WCAG and readability rules, and export the result without leaving the tab.

Problem

AI page generators are good at producing something plausible and bad at proving it works. Teams end up with layouts that look fine but have no accessibility or readability checks behind them.

Approach

I split the product into a generative side and a verification side. Blueprint generation uses Anthropic Claude first, with an Ollama fallback and a deterministic mock so the tool still works if the API is down. The audit engine is intentionally not an LLM; it's a 38-check pass over real HTML using Cheerio, WCAG 2.1 luminance math, and Flesch-Kincaid scoring so findings are repeatable.

Project method

Detailed process

Requirements

The requirement was to reduce idea-to-page time without accepting the usual weakness of AI page generators: plausible output with no proof of quality. The workflow therefore had to capture a brief, produce a structured blueprint, generate theme tokens, audit the result deterministically, and export usable artifacts.

Stack choices

The app became a Next.js monorepo so the studio UI, agent runtime, theme engine, UX audit package, and shared schemas could evolve independently. Provider calls were abstracted behind Anthropic, OpenAI/Gemini, and Ollama fallbacks, while Cheerio and the audit package handled HTML inspection without relying on model judgment.

Design system

The UI follows Brad Frost's atomic design: atoms, molecules, organisms, and templates with imports flowing downward. The visual system uses Product Studio tokens such as ps-canvas, ps-surface, ps-raised, ps-border, ps-accent, ps-ink, and severity colors, which made the studio feel like a coherent workbench rather than a collection of generated panels.

MVP definition

The MVP was the four-stage path from brief to blueprint, audit, and export. Multi-page generation, CMS integration, collaboration, and analytics were explicitly kept out so the first version could prove the core promise: generated page structure plus measurable UX quality.

Progression

Development started with canonical schemas for Brief, PageBlueprint, ThemeTokens, AuditReport, and FixPlan. The UI then layered in review checkpoints, run history, command palette actions, preview viewports, and export formats once those contracts were stable enough to keep the generated artifacts predictable.

Takeaway

Solution

Four stages — brief, blueprint, audit, export — take a product description to a scored page spec. A command palette keeps actions close, and exports to JSON, CSS tokens, full HTML, or a client-side ZIP without a server round-trip.

Evidence

Highlights

04 signals
  • 0138-check deterministic audit engine, independent of any LLM
  • 02Three-provider fallback chain (Claude → OpenAI/Gemini → Ollama) so generation does not hard-fail
  • 03Live sandboxed preview pane with desktop / tablet / mobile viewports
  • 04Fully client-side export pipeline — no server round-trip on download

Screenshots

Open channel

Need a sharper frontend path?

Architecture, AI workflows, design systems, and interfaces that need to hold up after launch.