AI-Native Product Design Lab

The future of healthcare is disease-first, brand-second. AI Patient Support is the working product that proves it. Instead of patients searching across twelve brand sites for assistance, they search by their condition and find every option in one place: support programs, savings paths, studies, videos, and patient stories, verified and organized by condition rather than by the company selling the drug.

Every drug page pulls from the public record: medication information from DailyMed and openFDA, trials from ClinicalTrials.gov, with the government record always one click away and a plain statement when something cannot be traced to a source. Media is tagged by length, source, and category so a patient can find a two-minute video or a full study without wading through either.

The lab now ships. CourtVisual is live at courtvisual.com: every game scored 0 to 10 for what is worth watching, with live fixtures, standings, and broadcast data behind it, and a Supabase backend with row-level security. AI Patient Support is live at aipatientsupport.com. ClinicalTrialsForMe is in preview at clinicaltrialsforme.com, with the architecture complete and the cancer trial path loaded end to end as the demonstrated vertical.

Three products, one conviction: decisions deserve current truth, rendered plain enough to act on. Each one takes live, authoritative sources and compresses them into a display a person can act on in seconds.

An AI-assisted dashboard for media buying and mix decision support, built in v0. Channel performance, engagement metrics, and AI-generated insights and recommendations: “Increase budget allocation to social media campaigns by 15%,” “Consider reducing radio ad spend and reallocating to higher-performing channels.”

Built and refined over 3 days. Would have taken a 4-week sprint in 2022, with a team of designers and engineers.

Design-system governance only works if drift is visible. The fastest way a client’s system erodes inside an agency or vendor workflow is the detached instance: a component pulled off the system, modified, and silently disconnected from updates. Figma flags none of this. The layer name turns from purple to black, and the debt accumulates invisibly.

I built Detached Instance Finder to make that debt visible in seconds. The plugin scans a page or entire file and surfaces likely detaches using two complementary signals. Name match catches the easy case: a frame named like a component, since detaches keep their original name by default. Orphan is the harder one. It flags a plain frame sitting among instance siblings, the one black layer in a row of purple. That second signal catches detaches even after they have been renamed, which name-matching alone cannot. Built with the Figma Plugin API and JavaScript, published to the Figma Community.

What I find interesting as a design leader: the tell designers use by eye (purple versus black) and the thing the Figma API actually exposes (node type) are the same underlying fact. The plugin does not invent a new method. It makes the invisible visible, so governance becomes something a system owner can enforce, not just preach. This is the same instinct I bring to multi-brand system work at scale.

Three Custom GPTs, all still live in ChatGPT. The first was built the night OpenAI opened Custom GPTs to the public. UX Research Advisor advises on UX research and marketing opportunities using user-provided data; Product Story, Strategy and Case Study Partner generates detailed product definitions and case studies; AEM Design Assistant covers best practices for AEM, Adobe Target, and DAM for storytelling and personalization.

They are the dated record of early adoption, not the current practice. The practice now runs on Claude Code and agentic pipelines and ships production software rather than assistants. Anyone can still click through and use these right now.

This pattern did not start with AI. In 2018, before “conversational design” was a discipline anyone hired for, I built a voice-first medication-adherence prototype as an Alexa Skill on the Amazon Echo. Adherence was already a multi-billion-dollar problem, and every existing solution asked the patient to do the one thing they were already failing at: remember to look at a screen. The question was what a reminder looks like if you never have to look at it, if you can just ask and be answered in the room where you live.

I started where voice forces you to start, with the conversation and not the device. I wrote the directed dialog prompts, mapped every intent, utterance, and slot, and drew the full conversation tree by hand before any prototype existed. The chart below is that architecture: user utterance to identified intent, to conditions of response, to device-specific response, to follow-up. Then I prototyped in SaySpring and tested with real people, designing the repair paths for when the skill heard the wrong thing and making sure it never left the user without a next move.

It is the same instinct the AI work runs on. The architecture around the interaction is the product, whether the surface is a voice skill in 2018 or an AI agent today: map the intent before the interface, design the repair before the success state, and treat what the system says as carefully as what it does. I build for new touchpoints early, on purpose, so that by the time a team needs them I have already learned where they break.