P PPROpportunity System Stage 2 review

Client briefing • August 2026

From a matching MVP to an operational opportunity system.

Stage 2 turns a promising scoring concept into a continuous, controlled workflow—bringing live property supply, operator requirements and explainable shortlists together.

Start the story

Live in production.
Deployed on two Hetzner servers since 28 August 2026.

01 — The journey

The MVP proved the idea.
Client feedback sharpened the product.

The decisive lesson: a high overall score is not the same as a commercially suitable match.

0129 MayMVP delivered

Core concept proven

022–19 JunClient review

Suitability rules sharpened

0319–23 JulStage 2 agreed

Operational scope defined

0423 AugBuild finished

Development complete

0528 AugDeployed live

Two-server production setup

Then

A one-direction scoring MVP

One requirement against a static Rightmove dataset, ranked mainly by a weighted overall score.

Now

A controlled opportunity engine

Continuous multi-source supply, matching in both directions and qualified shortlists that explain every decision.

02 — What Stage 2 delivered

Six changes that turned the concept into a product.

Each improvement directly addresses a limitation exposed by the MVP or enables the system to operate continuously.

01Better decisions

Commercial must-haves now decide qualification.

Five universal rules and nine operator-specific criteria are kept separate from ranking—so a strong average can no longer hide a decisive weakness.

14 criteria • clear pass/fail policy
02Human control

AI suggests the policy; the user approves it.

Every requirement receives a tailored recommendation. Users can review weights, mandatory rules and thresholds before the task starts, and that approved policy is frozen for repeatability.

Suggested → reviewed → versioned
03Two-way matching

Start from an operator—or from a property.

The original requirement-first journey remains, while the new property-first workflow reveals which eligible operators fit a particular opportunity.

Requirement ↔ Property
04Explainable results

Qualified shortlists are the default, not the exception.

Every result carries scores, confidence, evidence, gate outcomes and failed mandatory rules. Near misses are available deliberately, but never mixed into the normal shortlist.

Qualified-only review & export
05Live property supply

Rightmove became a service; LSH joined the pipeline.

Both managed sources now feed one source-independent property contract, with manual runs, daily scheduling, status, counts and safe failure reporting in a Control Room.

2 sources • one consistent pipeline
06Catalogue intelligence

The catalogue now understands change over time.

Stable identities preserve listing history, safe snapshots detect removals, and conservative grouping prevents the same physical opportunity appearing twice in matching.

Updates • removals • duplicates

03 — How a match works now

Factual eligibility first.
AI evidence second. Qualified shortlist last.

Step 1 • Protect the shortlist

Hard facts are checked before AI.

Availability Size Tenure Location

Known factual conflicts stop early, saving cost. Missing evidence is shown as uncertainty—not invented as a rejection.

Step 2 • Apply the approved policy
5 universal+9 tailored

Mandatory status decides whether a match is admissible. Weight decides how qualified matches are ranked.

Step 3 • Make the outcome explicit
QualifiedHard-gate rejectedMandatory rejectedAI scoring failed

Every pair reaches a clear state with evidence retained for audit and explanation.

04 — A major model upgrade

We moved from Gemini 3 Flash to GPT‑5.6 Luna.

The switch improved the quality and consistency of matching results while materially reducing model cost—a rare upgrade that strengthens both the product and its operating economics.

01Betterevidence-led outputs
02More consistentstructured decisions
03Lower costper AI workload

Prompts and default qualification thresholds were recalibrated alongside the model change—not simply swapped underneath the old workflow.

05 — One connected workflow

From source listing to qualified opportunity.

01

Ingest

Rightmove and LSH arrive through one controlled format.

02

Maintain

Updates, removals and likely duplicates are handled safely.

03

Normalize

Property facts and operator needs become comparable evidence.

04

Qualify

Facts are gated first; approved rules then decide eligibility.

05

Shortlist

Only qualified, explainable opportunities surface by default.

06 — The transformation at a glance

MVP → current product

CapabilityMVPStage 2
Property supplyStatic Rightmove datasetContinuous Rightmove + LSH
Matching directionRequirement → propertiesRequirement ↔ property
Decision policyCommon/static weightingTailored, reviewed and frozen
EligibilityStrong scores could compensateMust-haves independently qualify
Result reviewBroad score-led resultsQualified-only shortlist + near misses
Listing lifecycleReimported datasetStable updates, removals and grouping
OperationsNo source control layerControl Room + UK daily schedule
AI engineGemini 3 FlashGPT‑5.6 Luna

07 — Where we are now

The build is finished, and it's live in production.

Stage 2 is deployed on two dedicated Hetzner servers — one running the main application and database, one running the scrapers. Rightmove and LSH are both live, connected sources.

Product capabilitiesImplemented
Automated verificationStrong coverage
Production deploymentLive — two Hetzner servers
Managed sources runningLive — Rightmove + LSH