Buzzvil already knew how to reach people likely to be interested. I built the Interaction Composer to explore the next question: how could we turn that knowledge into a different brand experience for each audience?
One everyday wash. Start with the moment you want to make your own.
Choose a bottle to begin with, or a refill for the one you already keep.
The same shell, with a different way to wear it. Start with your route.
Keep it simple with a tee, or add an overshirt. Try both combinations.
From 2025, Buzzvil’s focus expanded beyond getting a campaign seen or clicked. Our targeting and retargeting, including reaching people who had already shown interest, gave us a strong starting point. The next challenge was to shape what those people experienced with the brand.
What we know about a group’s interests and behavior, informed by Buzzvil’s data and analysis.
The brand’s identity, product information and insights, so the interaction belongs to that brand.
The action the campaign aims to encourage, such as a purchase. This is its key performance indicator, or KPI.
Choose and combine interactions
for this audience, brand and goal.
A user segment is a group of people with shared characteristics. The design challenge was to turn what we knew about each group into a relevant sequence of interactions, while keeping the brand recognizable and the advertiser’s goal clear. One generic landing page could not express all of those differences.
My proposal was to give the AI a library of interaction patterns it could combine using audience data, brand context and the advertiser’s goal. Alongside the product team’s research, I built the Composer and a testing tool, the UI Cooker, to assemble and compare those variations.
Less.✓ Keep the bottle.
Refill the ritual.
A choice. Immediate feedback.
Let a gesture uncover the reward.
Build anticipation around a reveal.
KEEP Discover a preference, one card at a time.
Find the things that belong together.
✓✓Turn product recognition into play.
YOUR PICKA quiet essential.
Made to begin again.
“Finally, a bottleMeet your everyday
I keep.”
Narrow the choice. Add a reason to trust.
id: 'CouponCard',
intent: 'reward',
flowPredecessors: ['FlipCard', 'ScratchReveal', 'SpinWheel'],
compositionRules: [{
role: 'reward',
constraints:
'Use as the reward reveal after a discovery interaction. ' +
'Always pair with a CTA so users can act on the coupon.'
}]Original manifest excerpt. Each component carried design intent: where it belonged, how it behaved and what came next. Recipes combined these parts into flows.
Generating a flow was only the first test: people also had to stay with it. The Composer ran in a few campaigns, where usage data drove dozens of refinements. Here is one focused example, addressing people leaving before reaching the product.
I kept the reward visible throughout the quiz, so every answer would show progress toward a clear payoff. Compare the two versions, then try the three questions.
The reward stays in sight. Every answer makes progress tangible.
Good habits.Your reward is ready.
Explore the Daily Wash.
Three small discoveries. One daily essential.
Reconstructed refinement from campaign feedback. Its effect on conversion remains unconfirmed.
Alongside my Composer experiments, the product team tested instructions and supporting tools for the AI, without restricting it to a library of interface components. Results were stronger and more flexible. Generating and validating an experience for each audience group on the server made the extra generation time acceptable.
One layer.
No set route.
Judge the experience before anyone sees it.
Awaiting review Audience: new visitors
Shared components worked well for the possibility of live generation: a new experience for each session.
Brand-research tools and design guidance gave the model more freedom. Admin-side generation gave us time to review each variation.
The campaign platform serves a validated interaction to each user segment. If a segment underperforms, we can generate and validate another version on the server. Generation time stays outside the user’s experience.
Our UI kit made generation fast, but limited the quality and range of the results. Matching the model’s freedom would mean a much larger library and a growing system to maintain. The reason to build it became less convincing.
Brand details, user data and the advertiser’s goal proved more valuable inputs than a catalog of components. A clear prompt connected them. Richer context gave the model room to produce stronger experiences, which we validated before delivery.
As models improve, we’re gradually removing layers that earlier models needed. What remains is the essential context: user data, the brand, the goal and a clear prompt connecting those factors to the goal.
The direction, simplified. The prompt logic remains proprietary; human validation still happens before delivery.