An AI-assisted peer-accountability platform built and delivered for European life-sciences CEOs entering the US market.
Prototyping & Technology Lead
Spring + Summer 2026
~20 weeks
4 designers + CEO XB co-founders
Figma + Figma MCP
Claude Code
React / Tailwind
Vercel + Anthropic API
Maze
A confidential cohort of 10 to 12 CEOs is the app’s home: your peers, your facilitator, your next meeting
A monthly business update drafted by AI that reads your own history and your cohort’s activity, and never posts without your approval
Every feature works toward the same goal: building real connections inside your cohort, starting with a 1:1 coffee chat
A new member finishes onboarding and books their first coffee chat with zero help
Logging a monthly update feels like a five-minute win, not a chore
A live product the client can put in front of a CEO with no walkthrough needed
Breaking into the US market with no roadmap and no local network is hard. It’s also lonely. CEO XB’s cohort programs fix that in person, and our brief was to carry the same connection into the digital platform, so members stay connected between meetings.
CEO XB’s 2025 platform was built around an 8-module Mini MBA. Our 2026 brief was the next chapter: peer accountability, execution, and connection between the CEOs themselves, validated in a real product.
I built and deployed a React web app, live on Vercel, that repositions the platform around the cohort: onboarding through approval, a cohort-first home, a six-screen coffee flow, and a monthly update drafted by AI through the Anthropic API, grounded in the member’s own history. Then I rebuilt the Figma file to match the deployed code, frame for frame.
A walkthrough of the live product: Julie’s first day in her cohort, the coffee chat that starts it, and the AI-drafted monthly update that keeps her coming back.
Rapid Prototyping (React), AI Feature Design & Integration, Design System Reconciliation, Usability Research, Figma MCP Automation, Technical Stack Decisions
The C Suite: Laura Cochran (Product Design Lead), Somya Bhatia (Client Liaison), Emily Yang (Product Designer), and myself. Clients: the CEO XB co-founders. I owned the coded build, the Vercel deployment pipeline, and the Anthropic API integration.
Spring 2026: concept testing and definition. Summer 2026: the coded build, a moderated usability study, and the final handoff.
An evaluative study designed around the client’s own words: easy, intuitive, not a chore.
A moderated think-aloud study through Maze: three core tasks, 100% task completion. The findings came from watching where people hesitated along the way.
38% of participants entered the flow through something else first. “Honestly, that doesn’t feel like a button. It just feels like user info.” An affordance failure, fixable at the component level.
63% of participants called out a noisy home screen. “There’s too many things that all need my attention. What do I do first?” The hierarchy worked. It was pointing at the wrong thing.
63% of participants didn’t understand the label; some read it as physically relocating the company. Inside the flow, the concept landed. The study’s lowest ease-of-use rating traced to a word, not a workflow.
75% of participants preferred to write their own update and have AI improve it. They also expected Draft to read their real history: “I expected draft for me to gather data from across my account.” It didn’t at the time, so I built it in.
Every finding became a live change: a review-and-confirm state before posting, an AI restructured from generator to grounded editor, last month’s update anchoring the draft, and the confusing label renamed. Every change the client requested from the study shipped in the final build.
We mapped every screen of the 2025 build into one IA artifact: five tabs, a course spine, a static member directory. The map made the problem legible before anyone designed anything.
We tested six concept prototypes against a competitive audit spanning Vistage to Duolingo. What survived became the design vocabulary: activity visibility, time-pressure, a warm first coffee.
I designed the stakeholder workshop that turned six concepts into one structure. The client’s number-one ask: the cohort page IS the home. The revised IA marked every 2025 feature kept, demoted, or retired, so nothing disappeared silently.
I stood up a React/Tailwind app with auto-deploy to Vercel and rebuilt the team’s Figma flows as working software: 18 onboarding states, a cohort home, six coffee screens, a 12-screen update flow, and a progressive web app (PWA) layer that feels native on an iPhone.
Fixes from the usability study were deployed straight to the live build. Then I rebuilt the design file to match the deployed code, so handoff has one source of truth.
A cohort-first app where onboarding ends in a scheduled first coffee, the home screen is your cohort, and every surface offers a path to a 1:1.
A grounded AI update mechanic: Draft, Improve, and a cohort-matching Suggest, all running on the Anthropic API and reading the member’s profile, last month’s update, and the cohort’s activity.
A six-screen coffee flow with an availability-limited calendar and a personal message step, so scheduling a 1:1 takes under a minute.
Code and Figma that agree: a final-screens file rebuilt from the deployed build, every frame on live design system components.
We inherited the design system from the 2025 build. I reconciled what existed, extracting the token set into the codebase and relinking local copies to their library sources, then updated the system for 2026, including new iOS 26 native components, so it lines up with contemporary iOS design systems.
The brand was locked: Midnight navy, the 2025 logo, the existing palette. The craft was making a locked language feel native in 2026: true iOS glass treatments extracted from component specs, real status bar vectors, and one WebKit bug fixed, iOS under-reporting the web app’s viewport by 62px. Details like that are why the product reads as an app, not a webpage.
The first screen of onboarding says it plainly: log a monthly business update in seconds, have 1:1 coffee chats with CEOs worth meeting, and events that keep the connection going. A new member understands the app before creating an account.
Julie opens the app to her own cohort: the peers she’ll spend the program with, and one recommendation, a CEO she hasn’t met yet with a button to schedule a coffee. Home is built for connecting with her cohort, so the first thing the app asks her to do is meet one of them.
Julie picks a date from Sid’s real availability, shown in her own time zone. The chat is 20 minutes, and the app sends a suggested topic the day before, so neither has to prepare anything.
One clear message: you’re on Sid’s calendar. The screen spells out what happens next, a confirmation when he accepts and a video link 24 hours before the meeting, so there’s nothing to track and nothing to wonder about.
The AI read Julie’s July update, saw she focused on Fundraising and Operations, and drafted August’s starting point by asking what moved forward since. One tap improves what she wrote. She edits and approves everything before it posts; the AI never publishes on its own.
Posted updates land in the cohort’s shared feed: each peer’s win, whether their numbers are on track, and what help they’re looking for. Julie can see which peers face the same challenges she does, and that’s who she grabs coffee with next.
An AI “Draft for me” button that generated plausible text from nothing, on a home screen where a recommendation card, a due card, and a join button all shouted at once.
A monthly update that opens with the member’s biggest win of the month, color-coded KPI status, an AI that edits and cites where its text came from, a review-and-confirm step before anything posts, and a home screen that points at the coffee chat.
The deployed build runs the full journey, and the AI features run on the Anthropic API. Add it to your home screen on iPhone and it behaves like a native app.
Open the live build ↗When the code and the design file disagreed, the code won, and I updated the file to match. A handoff needs exactly one source of truth.
Three separate bugs came from the same cause: content that appeared on two screens was built twice, and the copies drifted apart. Now anything that appears in more than one place is built once, as a shared component.
Participants assumed the AI read their real data, and trusted it less when it clearly didn’t. So I gave it real data, an editor role instead of a ghostwriter role, and provenance on every output. An AI feature earns trust by showing its sources, not its confidence.
We tested with professional proxies, not European life-sciences CEOs. The navigation and affordance findings transfer. The premise findings don’t, and I’d trade two weeks of polish for more sessions with real CEOs.
The bet is that a posted update leads to a coffee chat. We built both halves and validated each separately. The connective tissue, seeing a peer’s update and acting on it, never faced a participant.
A deployed product can collect behavioral data passively, and ours mostly didn’t. The next version ships with analytics on day one, so decisions about the platform come from how CEOs actually use it.