I build AI products at the intersection of emerging technology, customer problems, and measurable outcomes.

I had a wardrobe full of clothes but still struggled to decide what to wear. Most wardrobe apps treat clothing as inventory rather than understanding the person, their preferences and context.

• —Wardrobe digitization, so a real closet becomes something a system can reason about.
• —Image understanding to describe garments in terms a styling decision can use.
• —A personal style profile that holds preference, not just category and colour.
• —Contextual outfit generation across occasion, weather and mood.

The hard question is subjectivity: how do you get an AI system to understand something as personal as style? I treated taste as something to be learned from behaviour rather than declared in a settings screen, and designed the surface so every recommendation is easy to reject cheaply — because rejection is the highest quality signal available.

Personalization only feels intelligent when the system is visibly willing to be wrong and adjust. Confidence without correction reads as a stranger telling you what to wear.

Deciding what to eat is a small decision made every single day, and the cost is not the cooking — it is the deliberation. Recipe apps answer a search query; they do not remove the decision.

• —A constraint model covering preferences, ingredients on hand, dietary limits and time available.
• —Suggestions that commit to an answer instead of returning a list to browse.
• —Conversational refinement for the second-guessing that always follows the first suggestion.
• —Variety tracking, so the assistant does not quietly converge on the same four meals.

Decision fatigue is a UX problem before it is a model problem. The product's job is to narrow, not to enumerate — so the default surface is one recommendation with a reason, and everything else is available only if asked for.

Users trust a confident single answer more than a ranked list, provided the reasoning is one line long and easy to argue with.

Most AI storytelling collapses into a chat window. The interaction model is generic, so the experience feels like prompting a model rather than being inside a story.

• —An authored interaction model where the story, not the chat transcript, is the interface.
• —A division of control: the system owns continuity and pacing, the reader owns direction.
• —Narrative state carried between scenes so coherence survives long sessions.
• —Deliberate constraint on user input — fewer, better choices instead of an open text box everywhere.

The design question was authorship. Give the user everything and coherence collapses; give them nothing and it is a novel. I drew the line at direction versus continuity, and let the interface make that boundary obvious rather than explaining it.

Interaction design does more for the feeling of magic than model quality does. The same generation inside a considered surface reads as a world; inside a chat bubble it reads as output.