From AI experiment to product system.

Engineering was already researching models. Leadership had ideas about what might be valuable. What we did not have was a finished brief, a defined interaction model, or even a clear answer to where AI belonged in Metaimpact.

I started broad: onboarding, data, profiles, dashboards, recommendations, content creation, network questions. The sketches were less about choosing a UI and more about mapping where intelligence could remove friction.

When I saw the first implementation in motion, Beacon was becoming too visually dominant. That was fine when the conversation itself was the task—but terrible when a user needed to read or edit the product underneath it.

I pushed for a smaller global entry point, then designed a dockable interface that could open as a focused surface, sit beside the work, resize, undock, and coexist with other panels already using the rail.

A prospective client had a heavily manual process for building internal project estimates. Multiple contributors and departments fed information into the same templated Word document, while contractor choices in one stage could affect budget, timing and decisions elsewhere.

The ambitious vision was scenario analysis: pull in contractor estimates, model tradeoffs, and recommend viable plans. Great idea. Too much for an MVP.

We narrowed the first step to something useful and reusable: a process-aware agent that could guide contributors through intake and synthesize the result into a cohesive document.

As we learned what agents actually needed, the profile evolved beyond discovery. Creators needed to control system instructions, knowledge sources, feature access, customization and performance visibility.

This work happened through several rounds with engineering—balancing what the system required with an experience that remained understandable, while weaving in shared platform features such as Properties.

Early Beacon work required a lot of Figma because the fundamental interaction model was still unknown. Today, the UI is established enough that I can spend more time defining how a new problem should work across the system.

I map the problem, identify the context the agent needs, define how generated content, references, attachments and follow-up actions fit together, then give engineering a pattern that can be reused the next time something similar appears.

Beacon started as an emerging-technology experiment with no settled product model. Over roughly a year, I helped turn it into a set of repeatable product patterns: a flexible global assistant, specialized agents, discovery and activation, agent configuration, proactive concepts, workflow-specific prototypes, contextual entry states and embedded AI behaviors.

Not every idea shipped, and that is part of the story. The work progressively clarified where conversational AI was useful, where purpose-built interactions were better, and where the architecture itself had to change before the product could deliver on the bigger vision.

That discovery now continues in Hubs and Pages, where I’m applying the same lessons to an architecture designed for deeper human-in-the-loop AI creation.