Building a teammate instead of buying a tool

Geoforge is an agent-first GIS platform that runs consultant-grade power diligence on any US site. Questions that conventionally take teams weeks come back in hours: whether a parcel can draw enough power, how long interconnection will take, and what sits ahead of it in the queue.

Speed is what the race to powered land rewards. A developer needs to reject bad sites fast enough in screening so that the good ones can be pursued before someone else reaches them. Geoforge delivers that speed by pairing forward-deployed engineers with AI agents, so what a design partner asks for on one call can be in the product by the next.

Holding that pace with the teams sitting 9 time zones apart challenged us to think creatively about how we can continue to scale.

What we needed was a generalist who could interpret what a customer wanted, turn it into a ticket, build the fix, and confirm it the way a customer would. So we built one.

Woz is an AI agent we customized to support Geoforge's forward-deployed engineers. Woz has an email address at the company domain, sits in Slack where it can be tagged by anyone, and executes Linear tickets delegated to it end to end.

The first time we used Woz, we were on a call with a customer who described a result that looked wrong. We tagged Woz in Slack and it began triaging while the call continued. Before the call ended, the fix was deployed and the customer was delighted to confirm it by refreshing the page.

But setting up Woz as a coding agent was the easy part. What makes Woz unique is that it is trained to use Geoforge's own product. A sandboxed browser lets it reproduce the bug directly rather than reason about it from the code. That is what makes Woz useful as a support engineer and an FDE rather than only as a software engineer.

Build vs. buy. Nothing on the market could be self-hosted with direct access to Geoforge's product, so Woz was built rather than bought. We see new AI coworker products launching constantly, and Geoforge has kept Woz for two reasons:

• One memory, persisted context. In most AI coworker products, context does not carry across connected tools. A conversation that happened in Slack is not available when the work moves to GitHub. Woz runs as a single vertically scaled server serving one company, so every session lands in the same place. Vendors building for thousands of customers decouple their architecture to scale horizontally, and you notice that in how fragmented the memory feels.
• Cheaper to run than alternatives. Building on an open-source harness makes the underlying model provider swappable. Geoforge can test new open-source models as they land, use subscription-based plans, and control its own inference costs, none of which is possible on a per-seat product.