Project notes

Summary

Garden subdivision / small residential infill was chosen as the first MVP template because it can be tested with mostly open geospatial data and a clear set of measurable rules.

• High Wycombe has many suburban residential plots with rear gardens, which gives the model enough candidate supply to test against.
• The basic inputs are available without paid data: parcel polygons, building footprints, roads, planning constraints and local sale prices.
• The opportunity can be expressed as a spatial problem: find residential parcels with enough remaining rear-garden land, plausible access and no obvious hard constraint.
• The output is easy to review on a map, so false positives can be checked quickly by a person.
• It is narrow enough for an MVP but still representative of the broader product idea: combining land, planning and value signals into a ranked shortlist.

Why High Wycombe

High Wycombe was chosen because it has a useful mix of property stock and data availability for this first test.

• Buckinghamshire is moving more planning data into standardised open-data formats through the national planning data programme.
• The area has better-than-average prospects for joining local planning, constraint and property datasets without starting from a fully closed data environment.
• High Wycombe has many suburban plots with rear gardens, which matches the selected opportunity template.
• The town has enough residential density and transaction history to test the value and ranking logic.
• The council planning systems are in transition, so planning-history coverage is not perfect yet. The benefit is that the direction of travel is towards more open and standardised data.

Current scope

• Location: High Wycombe, Buckinghamshire.
• Opportunity type: garden subdivision / small residential infill.
• Status: MVP built quickly to test whether the data pipeline and map workflow are viable.
• Not included yet: other opportunity templates such as garage replacement, backland access strips, edge-of-settlement land or brownfield sites.

What the app shows

• A map of candidate parcels.
• A ranked list of possible sites.
• Basic filters for confidence score, garden size, access type, saved candidates and text search.
• A site panel showing garden area, frontage, approval signal, estimated GDV and a short rationale where available.
• Saved candidates in local browser storage only.

How it is built

• Frontend: Next.js.
• Map: MapLibre GL with a free basemap and candidate parcel overlays.
• Database: Supabase Postgres with PostGIS.
• ETL: Python scripts using geospatial libraries to load and prepare source data.
• API: Supabase PostgREST views and an RPC returning candidate GeoJSON.

Data sources

HM Land Registry INSPIRE Index Polygons

Indicative freehold parcel geometry.

OS OpenMap Local Buildings

Building footprints used to estimate remaining open land inside a parcel.

OS Open Roads / road geometry

Used to estimate road frontage and infer the front of a plot.

planning.data.gov.uk constraints

Used for constraint checks such as conservation areas, Article 4 directions, green belt, flood risk and related planning layers where available.

HM Land Registry Price Paid Data

Used as a rough local value signal from nearby historic sales.

Buckinghamshire planning history

Used, where available, to estimate comparable approval rates. Coverage still needs checking and cleaning.

Candidate logic

1. Load and clip source datasets to the High Wycombe area.
2. Start from parcel polygons.
3. Subtract building footprints from each parcel.
4. Estimate rear-garden area using road-facing frontage and a front-setback heuristic.
5. Measure parcel area, garden area, garden width, frontage and likely access type.
6. Check whether known planning constraints intersect the parcel.
7. Calculate a score from garden size, access, likely dwelling fit, approval signal and value signal.
8. Expose the results to the frontend as rows for the list and GeoJSON for the map.

Scoring fields

• Garden area: larger rear gardens score better.
• Fit: whether the remaining land appears large enough for a small dwelling and access.
• Access: street frontage is treated as stronger than unclear or backland access.
• Approval rate: local comparable planning outcomes where data is available.
• Value: nearby sale values used as a rough GDV input.

AI functionality

The core candidate selection is rule-based geospatial processing. AI is used around that process, not as the source of the parcel geometry or scoring.

• Candidate rationale: generate a short plain-English explanation for why a parcel scored the way it did.
• Planning text parsing: extract useful fields from planning descriptions, officer reports and decision notices where structured data is missing.
• Comparable site review: help group historic applications that look similar to the target opportunity type.
• Future image review: aerial imagery or LIDAR-derived outputs could be checked for things the vector data does not label well, such as garages, hardstanding, tree cover and usable garden shape.
• Human review support: produce site notes for a shortlist, but not make the decision on whether a site is viable.

In this MVP, the visible AI output is the rationale field shown on candidate cards where it has been generated. The ranking itself comes from the measured fields and scoring rules.

Known gaps

• The MVP only covers one town and one opportunity template.
• INSPIRE parcel boundaries are indicative and need checking against title documents.
• The rear-garden calculation is a heuristic and will be wrong on some plot shapes.
• Planning history data needs more cleaning before the approval signal should be trusted.
• Owner names are not included. They would require paid Land Registry title lookups on shortlisted sites.
• LIDAR, richer comparables and manual validation have not been fully built into the workflow yet.

Next useful improvements

• Manually audit the top candidates and record false positives.
• Add LIDAR height checks for dwellings, garages, trees and usable flat ground.
• Improve comparable planning history and label approved / refused examples.
• Add exportable site summaries for shortlisted parcels.
• Add more opportunity templates once this first one has been validated.