The products I find most interesting are the ones nobody has shipped before, and there’s no better time to be building them. At IBM AI Incubation, Flint went from brief to 500 users in twelve weeks, and Relay was prototyped while the agent framework it depended on was still being built. Building for emerging capabilities is one way to think about how to define what comes next. 

I design interfaces for systems that generate, reason and act. As models become more capable, the design problem is moving beyond prompting towards steering, delegation, supervision and intervention: deciding what humans should control, what agents should handle, and how the two stay coordinated over time. I’m interested in the moments when a new model capability appears and a new interaction model has to be invented and defined.

I run the first design team in IBM that submits PRs directly into the codebase. I work directly in code because the fastest way to understand a new interaction is often to build it. Using Figma MCP, React, Cursor, Codex, Claude Code and MCP workflows, I move between interaction design and implementation without treating handoff as a fixed boundary. That lets me prototype emerging capabilities quickly, test them in realistic conditions, and carry design intent much further into production.

Some of my work is designing the tools behind the product: model experimentation surfaces, research tooling, evaluation interfaces and internal dashboards. At IBM I’ve worked on AI-native patterns for Carbon v12 and tooling that helps teams inspect, configure and work with increasingly agentic systems and build more efficiently and faster. I’m interested in interfaces that make model behaviour observable enough to study, compare and improve.

Constraints often produce the most interesting AI design problems. On NatWest’s Cora+, the system had to avoid hallucinated financial advice; on a UK Government programme, generated output operated inside workflows with real legal consequences. My work in AGI strategy and technical AI safety has sharpened how I think about uncertainty, oversight, human agency and designing around what models should and shouldn't act.

The other half of frontier work is what happens when experimental systems meet reality. Cora+ served 11M+ NatWest customers, the UK Government work sat within a £400M programme, and Flint moved through continuous releases into client-facing use. At that scale, the design problem isn’t a perfect user following a clean flow. It’s making new capabilities survive legacy systems, compliance, edge cases and real organisations.