Five days, one pitch

The brief

R&D teams across multiple sites were working in silos. Formulation data in one system, experimental results in another, ML predictions nowhere near the people who needed them. Researchers were spending hours on work that should have been automated, and making decisions without the data that existed somewhere in the organization.

Dense brief. A few assumptions buried in it. Written by people who knew the chemistry deeply but weren't sure yet what the interface should look like. The kind where you can lose half the week going in the wrong direction.

Three agents, one file

Instead of working with AI as a general assistant, I set up three specialized agents and built a plugin to connect them directly to my Figma file. Each one had a specific lens. Depending on what decision I was making, I pulled in a different kind of pushback.

The Design System agent was the one that changed the output the most. I had loaded a variabilized Figma library into its context, so every component recommendation was already token-aligned. The file came out clean not because I cleaned it up at the end, but because the agent was enforcing the system throughout.

Skipping wireframes

I coded directly in the terminal, with AI feeding the iteration loop. The goal wasn't clean code, it was something you could actually interact with, not just click through. A working prototype makes a client feel the product. A static mockup makes them imagine it. For a pitch, that's the difference between "interesting" and "when can you start".

Once the prototype answered the right questions, I translated it back into Figma. Not the other way around.

That order matters more than it sounds. The usual design-to-dev handoff always loses something: a spacing decision that doesn't survive implementation, an interaction that gets simplified, a detail that felt obvious in Figma and gets lost in translation. Here it was reversed, I validated the build first, then fixed the UI layer. The Figma file became a faithful record of what actually worked, not an optimistic spec of what might.

Making predictions explainable

A prediction alone isn't enough. Researchers need to understand why. SHAP values break down every prediction into ingredient-level contributions, what pushes the result up, what pulls it down. This was the view that made the pitch click: it turned a black box into something a chemist could actually trust and act on.

The week

Five days, each with one clear output. Brief, setup, build, translate, present. The constraint forced a sequencing I wouldn't have chosen otherwise, and it was better.

What this changed for me

We didn't end up winning the pitch, the brief eventually shifted toward the data team's scope. But what stayed with me was realizing that the usual risk in a pitch, building something polished that misses the point, is actually a process problem. Starting with brief clarification removes it. Specialized agents enforce quality throughout rather than at the end. Prototyping in code instead of Figma makes the demo land differently in the room.

None of these things individually are new. Together, with a hard deadline forcing commitment to each one, they produced something I'm not sure I could have matched with three weeks and a traditional process.

This pitch is also where two essays published here were lived before being written: The handoff is dead. Now what? and The spec is the new wireframe.