Designing an AI that speaks for me

Recently, I took my hand-coded site and transformed it into a componentized, adaptable platform and even threw in a conversational experience as a fun experiment to build and shape called TonyBot.

👇 Skip the preamble and jump straight to the case study:

For decades now I have evolved and redesigned my portfolio site over and over again, building it by hand, learning, coding, and relearning HTML, CSS and JavaScript. I’ve always loved interactive design, so I taught myself front end development so I could tinker and build web experiences for clients and work.

Over the years as I focused more on design and UX, updating, redesigning and rebuilding my portfolio site became a regular opportunity to stay current and relevant in Front End. In fact, I designed and built my current site almost entirely in the browser.

More recently, I decided I wanted to take my site to a level I hadn't personally coded to before - an adaptable, componentized platform that I could easily build and tailor experiences and case studies. In the fall of 2026 I set out to modernize it with three goals: migrate it to a component system without losing its look, make sure it’s mirrored in a Figma design system, and add something new and fun to build, an AI assistant that answers recruiters’ and hiring managers’ questions about my career.

💼 Case:

Of course I’m not the first person or even designer to introduce a conversational experience on their site - but after shaping an entire Conversational Design practice at Ozmo and leading the effort behind Ozmo’s first shipped AI capability, I wanted to have a little fun building something that could work in an interesting way with my current job search.

Nor am I the first person to build or rebuild a site with Claude Code. I’ve been a daily Claude user for a while now and I still get excited about how fast you can translate an idea into something that works and functions.

⚠️ The problem:

Years of hand-coding had left my existing site’s styles layered and hard to change safely. Design and code had no shared system, so every update meant making the same decision twice. A static portfolio also asks a lot of a busy recruiter: the answer to their specific question is usually in there somewhere, but they have to dig for it.

An AI assistant could solve that last problem, but it creates a bigger one. An AI speaking for me in public is a trust problem. One invented metric or overstated claim costs more credibility than the assistant could ever earn back. This is where years of judgment matter: setting the foundational behaviors and governance that keep it accurate and consistent.

🛠️ Process:

Protect the look first

Before changing anything structural, I set up automated screenshot comparisons across every page at three screen sizes, plus checks for motion, keyboard use, layout shift and pages without JavaScript. Every change had to prove it didn’t break the experience before it could ship.

A system in code

The site moved to a component system built with Astro and React, keeping every public page at its existing URL. Colors, type, spacing and motion became design tokens in one source file, with colors named by role so a future dark mode is a matter of swapping values. Releases now flow from GitHub to a preview site first, and nothing goes live until I publish it. Before, I was literally dragging HTML files into a server folder in Filezilla.

The same system in Figma

I built a Figma library that mirrors the coded tokens, type styles and components, so design and code speak the same language.

Designing TonyBot

I made a few foundational decisions early:

Guardrails as design

I treated the assistant’s rules as part of the experience, and tested them with real questions. I wove in some specific guardrails like the following:

Reviewing the AI’s work was a lot like reviewing a team’s work. Several catches came from close reading, such as a draft that implied a cause-and-effect link my experience doesn’t support, and a concept that read as shipped, when it was really an exploration I use to show my process. Each one became a permanent rule and an automated test, including one that fails if the assistant ever mentions its own behind-the-scenes setup to a visitor.

Privacy by default

Conversations are preserved, but safely - with no IP addresses, contact details scrubbed automatically, and deletion after 90 days. A line under the input says so plainly.

The interaction

I call it TonyBot!

TonyBot wears a hat - just like I typically do as well. We’re almost like twins. I’ve always felt like a person who wears many hats. I love context switching between designing, researching, managing a team, leading a discovery, etc. I even helped TonyBot out with some quick animation in After Effects - he waves!

Again, my ultimate goal with this was building something fun, a conversational experience focused on my background. And with a componentized, adaptable site architecture, I can move it around easily and actively shape it. And while it needs to be accurate about my background, there’s also my tone and some playfulness to it.

I injected some fun questions and responses that reveal a bit about me - you can ask it about my dog Wilson for example. More importantly, I’ve shaped it to adapt to who it’s talking to. After its first answer, it asks who you are, whether a hiring manager, a designer or a recruiter, and adjusts its answers from there.

I also appreciate a conversational experience that can leverage more than just text responses. TonyBot will show you a video or graphic inline or point to the right resource on my site depending on each conversation.

✅ Results:

I now have an adaptable, componentized portfolio platform that has a conversational experience available for where and when I want to augment my content and experience stories. With a dialed in AI-enabled Design System I can easily build new experiences in minutes within my site using AI technology or I can jump into either code or Figma files for fine tuning.

Every release now passes 360+ automated screenshot comparisons, the chat adds no layout shift when pages load, and text contrast holds at 5.13:1 or better. Answers average about a tenth of a cent each.

The bigger result is a working model for leading AI the way I lead a team: set clear standards, review the output closely, turn every catch into a lasting rule, and own the quality of what ships.

Try TonyBot out:

Hi! Tony created this AI assistant to answer questions about his experience, projects and how he leads teams.
Pretty neat right? But not as cool as talking to Tony himself!

Your questionAsk

What did Tony lead at Ozmo?How does he build design teams?What's with the AI imagery?

Conversations are reviewed to improve this assistant and deleted after 90 days. Please don't share personal details.

Related Case Studies & Articles

• My ExperienceLeading teams

  AI Adoption

  with Psychological Safety

• My ExperienceOrchestrating modern

  Design Systems

  across Governance & Collaboration

• ArticleAI Convo-Design

  Convo Flow

  Conversational Design and effective communication.

Related Projects: