Sandbox

I design AI-native products by building them.

These are working tools I scoped, designed and shipped myself, from the interaction model to deployed code. Every one is live, and you can try it right here on the page.

  1. Good AgentAI UX pattern librarygood-agent.vercel.app
  2. FlowCraftVisual flow builder for AI coding agentsux-flow-builder.vercel.app
  3. VoiceForm LabResearch prototype for conversational AIaudio-visual-test.vercel.app
good-agent.vercel.applive

Good Agent

AI UX pattern library

73 interface patterns for agentic products, mapped to the nine moments of working with an agent, each with a working demo.

good-agent.vercel.app/patterns/ghost-completionOpen in new tab

Why I built it

A send button isn’t an agentic pattern. Teams designing for agents need patterns organised around the moment the user is in, not the component they want to build.

What I built

  1. 73 patterns mapped to 9 moments, from the first ask to living with what the system remembers, and grouped into 12 families
  2. A live, interactive demo for every pattern, not a screenshot
  3. A tradeoff and checklist for each moment, plus browsing by journey or by element with filtering
ux-flow-builder.vercel.applive

FlowCraft

Visual flow builder for AI coding agents

Design a user journey visually, then hand it to Claude Code or Cursor as structured context, not a screenshot to interpret.

ux-flow-builder.vercel.appOpen in new tab

Why I built it

AI coding agents can read screens and components, but not the journey between them: what triggers a transition, and how errors and decisions branch it.

What I built

  1. A canvas for typed screens, decisions, actions and errors, connected by the triggers between them
  2. AI flow generation with your own key for Gemini, OpenAI or Anthropic
  3. Typed JSON export that drops into a CLAUDE.md, a prompt or an MCP response
  4. Cloud sync on Supabase, open source under MIT

How I used AI

Designed and built by co-creating with Claude Code, the same kind of agent it feeds. The tool and the way it was made test the same idea.

audio-visual-test.vercel.applive

VoiceForm Lab

Research prototype for conversational AI

An instrument for exploring how visual and audio feedback shape the feel of talking to an AI.

audio-visual-test.vercel.appOpen in new tab

Why I built it

Conversational AI moves through states (idle, listening, thinking, speaking), and the feedback for each shapes how the conversation feels. Tuning that needs something you can play with.

What I built

  1. An audio-reactive particle visualisation for each conversational state
  2. A built-in synthesiser with oscillators, envelope, filter and delay to tune cues and timing
  3. Live controls for radius, speed, sensitivity and particle count