AI land sourcing across 42 local authorities in under two hours

01The challenge

Land sourcing was manual. A land manager applied broad filters on an off-the-shelf map tool, then inspected thousands of parcels by eye, cross-checking planning policy, flood risk and ownership. What each developer really wanted emerged through conversation and was never written down in a form a system could use, so nuance was lost and a search across the whole country was out of reach. Even a clear brief, such as sites near affluent schools but not too close to the main road, had to become a precise spatial query before software could act on it, and the same main road that suits an EV charging site counts against a housing one.

When we started, Virgin Land was in its infancy, with more ideas for technology than budget to build them. It needed systems that worked at its size then and could grow with it.

02What we did

1. 01Phase one — matching and letter sending automated on Google Sheets, Pipedrive and Zapier, so the team could test its model without paying for bespoke software
2. 02Phase two — after angel investment, data and operations moved to a custom-built Laravel platform, with Gigabyte acting as outsourced CTO
3. 032022–23 — client meetings transcribed and passed through an LLM that turns each developer's brief into structured rules for a land expert to approve
4. 042022–23 — every parcel tested against those rules with bulk PostGIS queries: about a million title polygons across two local authorities in under an hour
5. 052023–24 — scaled to 42 local authorities in under two hours by moving spatial joins into the database and fanning work out across AWS Lambda
6. 062023–24 — a reinforced LLM chain with few-shot examples and self-critique replaced regex rules, lifting requirement accuracy by 27 percentage points
7. 072023–24 — an agent that writes multi-step PostGIS queries from plain English, piloted on EV charging sites in Bristol

03What we built

The platform follows Verafind's process end to end: capture what a developer wants, find matching sites already in the database, source new ones, write to the landowners, share qualified sites through a client portal, and offer sites one developer declines to others in the network. Land managers see AI-generated matches, their own flags and each site's pipeline status in one place.

Experts stay in the loop. Every rule the agent extracts goes to a land expert to accept or reject with a reason, and those reasons refine the prompts. In the Bristol pilot, rejections such as a suggested site inside a conservation area fed back into the schema descriptions the agent reads and the way it reasons.

• Living requirements agent fed by meeting transcripts
• Automatic scoring of sites against each brief
• PostGIS parcel scoring at local-authority scale
• Plain-English queries across geospatial datasets
• Automated handwritten letters to landowners
• Client portal for developer review and feedback
• Reverse matching of declined sites to other developers

04The result

The first trials surfaced sites the land team had missed. By 2024 the platform was in daily use: the scoring engine covered 42 local authorities in a single run, and land managers worked from one screen of AI matches, their own flags and pipeline status instead of a separate spreadsheet macro. Matching the other way round, from one site to every developer's requirements, reached about 85% accuracy where the parcel data was rich.

The Bristol pilot also showed that how the database is described to the model, not the model itself, decides whether its queries are accurate. Along the way Virgin Land secured angel investment and was acquired by The Ringley Group, and Gigabyte remains Verafind's main technical lead.