Rowland Case Study

The Problem

Land departments at energy companies sit on decades of paper — leases, assignments, joint operating agreements, deeds, easements, division orders, revenue statements. That paper holds the answers the business runs on: who owns what, under what terms, expiring when, paying whom. Historically the only way to get those answers was to pay an experienced landman to read every page, and legal descriptions written in survey-system language rather than coordinates meant even a correctly extracted parcel still had to be mapped by hand.

Rowland was built to replace that with a pipeline. The product worked and the core ideas were sound, but by mid-2026 its engineering had been carried by a rotating in-house group — five names in five months, each handover costing weeks of ramp. Leadership was setting a roadmap on a monthly rhythm while delivery landed on a quarterly one. Underneath, the platform was unstable: documents stalled mid-pipeline or timed out with no signal why, deploys declared victory before rollout succeeded, and connection limits fought workflow concurrency settings so the system throttled itself under exactly the bulk loads customers cared about. Meanwhile, three missing features were stalling every enterprise conversation on the same objections — how do I verify this, can I hand this to my client, can my whole firm use it?

What We Did

We embedded two engineers directly in the team, shipping to production from the first week. The first fortnight went entirely on the floor rather than the ceiling: stuck and falsely-failed documents, deploy gating on real rollout status, orphaned Kubernetes namespaces quietly stealing work from live workers, worker slots aligned to the database pool, and production capacity raised threefold.

With the platform steady, we built the three features holding back revenue. Verification was the hardest and the most important — in title work, an extracted royalty fraction is worthless if a landman has to re-read the lease to believe it. We made Document AI the single source of truth for both text and page coordinates, which let us replace fuzzy text matching with deterministic click-to-locate highlighting: click any extracted value and the source page scrolls to and highlights the exact words behind it. The runsheet export turned Rowland from a tool that helps you do the work into one that produces the work product, complete with client branding and password-protected links back to source instruments. Multi-organisation support let firms, not just individuals, onboard.

We also built the GIS resolution engine and geospatial reference database that convert legal descriptions into real mapped boundaries using official cadastral datasets — PLSS, the Texas Original Land Survey, and county boundaries — and gave users page-level control over messy scans, so correcting one bad page reprocesses only what that page affects rather than paying for full re-transcription.

Impact

In nine weeks, a team 60% smaller than the one before it shipped 58 tickets and 75 merged pull requests — roughly 75% of the previous raw output, and about 1.7× as much per engineer. More telling than the volume was the shift in what the work was: three net-new customer-facing surfaces in the first nine weeks, against zero in the preceding quarter.

Rowland now runs in production on Azure Kubernetes serving enterprise land and energy customers, with a documented performance ceiling and a measured parallelization plan that has become the roadmap for its next phase of scale.