CASE NOTE / FOR REAL ESTATE AGENTS
Place-first relocation guides that capture leads, and a review step so an agent only sees signals a person has confirmed.
01 / CONTEXT
The build was not only a website. It connected static relocation guides, lead forms, scoring rules, analytics events, public records, and extracted signals into a portal that could support actual follow-up.
Two Southwest Virginia regions were in scope, and they were modelled as genuinely separate places rather than one market with a filter on top. The New River Valley covers Montgomery, Radford City, Pulaski and Floyd. The Roanoke Metro Area covers Roanoke County, Roanoke City and Salem. Each got its own towns, its own accent colour and its own standalone site, because a buyer relocating to Blacksburg is not reading the same page as a buyer relocating to Salem.
02 / WHAT SHIPPED
Place-first homepages, a library of topic guides, a cost-of-living calculator, a region-wide business directory, and an interactive map drawn from real boundary geometries with census-tract overlays. Per-area, subarea and local-resource pages go down to the town level, and every page carries a note about where its data came from and how fresh it is.
The sites are hand-written HTML, CSS and JavaScript emitted by a build command. That keeps them fast to serve and cheap to host, which matters when the whole point of the guides is to be found in search and shared in a text message.
The guide forms post to a self-hosted webhook rather than a third-party form service. Leads are scored as they arrive, normalised into a consistent record, and exported to a CSV shaped for the CRM the team already used.
A cookie-based visitor tracker and a funnel dashboard, built in rather than bolted on. No third-party trackers were required, so the behavioural signal feeding the lead score stayed in the same database as the leads themselves.
A web application with login and role scoping over the lead dashboards, guide readiness, guide content, and previews that do not publish. Signals surface here for a person to accept or discard; nothing reaches an agent unreviewed.
03 / THE PIPELINE
The interesting engineering is not the scraping. It is everything between a raw public record and something an agent should act on.
Public hiring and press sources are pulled in, then structured by a language model rather than by brittle per-site parsers, because the sources change layout constantly and a parser that breaks silently is worse than one that never existed. Each record is then checked against county property-ownership records, filled in from public professional directories, and matched to a person, which is the step that decides whether two similarly named records are one person or two.
Only after that does anything become a ranked signal, and even then a person reviews it before it is treated as a lead. That ordering is deliberate. A signal that is wrong about who someone is does not waste a minute of an agent's time; it costs them a relationship.
The test suite records real HTTP interactions as fixtures and replays them, so the pipeline can be tested end to end without touching a live source. It also means a source changing its markup shows up as a failing test rather than as a quiet gap in the data.
04 / WHAT MADE IT HARD
Regions had to stay genuinely separate in the output while sharing every importer, model and build path. The lead pipeline's market attribute is a different concept from the site region, and conflating the two would have quietly mixed one region's leads into the other's reporting.
Names collide, records are stale, and county property data uses its own conventions. Matching a record to the wrong person does not produce an obviously broken system — it produces a confident one that is wrong, which is far more costly.
Public data ages at different rates. Rather than hide that, every guide page states its sources and how current they are, so a reader can judge the number instead of trusting it blindly.
05 / PATTERN
This is the data-work shape Seam likes: take scattered public and first-party signals, build a usable workflow around them, and keep the final system plain enough that a small team can work it every week. The parts of this build that survived into Seam's own products were not the clever ones. They were the review queue, the provenance notes, and the decision to make a person confirm a signal before it became a lead.
06 / MORE FROM SEAM
A working system for generating weekly technician schedules across clinics, reviewing draft runs, filling gaps, and keeping an audit trail of schedule decisions.
Practical AI workflow consulting for businesses that need clearer processes, cleaner data, safer automation, or a useful internal tool.