Project Orbit

Agency work is coordination work. Pipeline lived in one tool, delivery in another, performance in a spreadsheet rebuilt every month. Nobody could answer a simple question, which accounts are quietly slipping, without an afternoon of assembly.

The brief mentioned AI. The real problem was that the underlying record of the business did not exist in one place, and no model can reason over information that was never joined.

We spent the first phase defining the objects the business actually runs on: accounts, opportunities, engagements, and the signals attached to each. Only once those were stable did we introduce retrieval and generation.

Every intelligent surface was then scoped to a decision with an owner and a deadline. Which leads deserve attention this week. Which engagement is drifting. What changed since the last review.

Nothing the system suggests arrives without its sources. Each recommendation shows the records behind it, invites correction, and records what the user decided, which is also how the system gets better.

The interface is dark, dense, and fast by design. This is a tool people keep open all day, so latency and information density were treated as experience decisions, not engineering details.