Protecting public land with sovereign AI

Work in progress — an early demonstrator we’re sharing to show where we’re headed.

Every council in NSW owns land it struggles to keep an eye on: reserves, road reserves, drainage corridors, bushland. Over time that land quietly disappears at the edges — a shed over the boundary, a back fence pushed into the reserve, a driveway sealed across the road reserve, fill and green waste dumped on public ground, vegetation cleared without consent. Individually small; collectively a steady erosion of public land, ratepayer value and environmental protection. Today it’s mostly found by chance, by complaint, or by a site visit years too late.

This is one of the problems we’re working on with councils right now. Land Encroachment is a new module on the Edgegenix platform, and below is an early look at how it comes together end to end.

See it in action

Walkthrough · a drone survey plays through the console: the map tracks the flight, the system pauses on each finding — shed, hardstand, fence, dumped material — and an on-screen assessment checks it against the council boundary. Near the end, a LiDAR view reveals a structure hidden under tree canopy.

Setting expectations: this is a concept demonstrator, not a finished product. The detections in the video are pre-baked and illustrative — they show the workflow, not a trained production model — and the LiDAR under-canopy view is a roadmap capability shown for illustration. We’d rather show you honestly where we are and where we’re going than oversell a demo.

From drone to decision

The pipeline · detection happens in the field; the boundary test, evidence chain and tenant isolation happen in the Sydney region; a person makes every call.

1. Survey. A drone (RTK/PPK where accuracy matters) flies the target area and captures high-resolution imagery.
2. Stream to the edge. Drone video can be streamed directly to an Edgegenix Edge AI device — a compact, ruggedised on-site compute unit that runs detection in the field, without a round trip to the cloud. That means near-real-time results, operation in areas with poor or no connectivity, and imagery that can stay on the customer’s own hardware.
3. Build the map. Imagery is stitched into a georeferenced orthomosaic (GDA2020 / MGA56) — the same coordinate framework councils and surveyors already use.
4. Detect. Our models identify what matters: structures and sheds, fences, sealed hardstand and driveways, dumped material, and clearing.
5. Test against the boundary. Each detection is intersected with the council’s cadastre and land register (community vs operational). This is the step that turns a picture into a finding.
6. Two confidences, never blended. We report detection confidence (are we sure it’s a shed?) and boundary confidence (are we sure it crosses the line?) separately, with the incursion margin. Cadastre isn’t survey-accurate, so we never pretend a boundary is exact.
7. Operator review. Every finding is advisory only. It triggers an investigation; a registered surveyor determines the boundary. The system issues no notices. Each decision is written to a tamper-evident evidence chain.

Because a camera can’t see through tree canopy, our roadmap adds LiDAR — laser returns that penetrate foliage gaps to reveal structures hidden beneath the trees that photogrammetry alone would miss.

What this is worth to a council

The value case is straightforward, and it compounds.

• Coverage that was never possible on foot. A single drone flight surveys hundreds of hectares in a morning — every boundary on the flight path checked against the cadastre, systematically, not by chance. The realistic alternative isn’t a cheaper survey; it’s no survey at all, plus complaints.
• Findings while they’re cheap to fix. An encroachment caught early is a letter and a conversation. Caught a decade later it’s a legal dispute, a survey bill, and at worst an adverse-possession claim on public land. Early detection moves the whole caseload to the cheap end of that curve.
• Recovered value, not just enforcement. Many encroachments end in regularisation — a licence or lease on proper terms — rather than removal. Findings open a revenue conversation as often as a compliance one.
• Officer hours spent on cases, not paddocks. Every advisory finding arrives as an evidence pack — imagery, coordinates, overlap area, both confidences — so staff investigate specific cases instead of walking boundaries hoping to spot one.
• A record that defends itself. When a decision is questioned — by a ratepayer, a councillor, or a GIPA request — the evidence chain answers for it.

To make that concrete with deliberately conservative, illustrative numbers: a council holding a few hundred parcels of community and operational land, flying its high-risk corridors twice a year, would review more boundaries in a fortnight of flights than ad-hoc inspection reaches in years — and a single regularised licence, or one legal dispute avoided, can fund a season of surveys. Every council’s numbers are different, which is exactly the modelling we’d do with you in a pilot, on your figures.

Who this is for

Property and open-space managers who answer for the land register. Regulatory and compliance officers who need evidence before they act. GIS teams who want findings delivered in GDA2020, ready to drop into the systems they already run. And executives who have to defend both the budget and the decisions made with it. If one of those is your desk, this was built for you.

Sovereign AI, by design

Here is what makes this defensible for government: the intelligence runs sovereign.

Council land data — parcels, land registers, survey imagery, findings — is sensitive public-sector information governed by NSW privacy and access law (PPIPA, GIPA). It cannot be shipped to an offshore, black-box AI. Edgegenix is built so it never has to be:

• In-country, in-region. The platform and its AI run in Australia (AWS Sydney, ap-southeast-2). Your data stays onshore — and with the Edge AI option, sensitive imagery can be processed on-site and never leave the customer’s own hardware at all.
• Per-tenant isolation. Each council is its own tenant. One council’s parcels, imagery and findings are logically and cryptographically separated from every other’s — enforced on every request, not by convention.
• The customer’s own intelligence. The operational-intelligence layer that reasons over each survey is scoped to that tenant. No pooling of council data to train a shared model; no answers leaking across boundaries.
• Auditable and explainable. Every AI assessment shows its working — which parcel, what overlap area, which tolerance was applied, whether a permit was on file — and every operator action is logged to the evidence chain. Councils can defend a finding, and answer a GIPA request about it.
• Advisory by design. The AI recommends; people decide. For enforcement on public land, that is not a limitation — it is the only responsible posture.

Sovereign AI is not a feature we bolted on. It is the reason a public agency can put this in front of its ratepayers and its lawyers.

Where we are

Land Encroachment is an active work in progress. The workflow you see — survey → edge and cloud AI → cadastre intersection → advisory finding → operator review, all on a sovereign, per-council tenant — is the target we’re building toward with council partners now. We’re sharing it early because the problem is real and immediate, and because the councils we talk to would rather help shape the solution than wait for a finished one.

If you manage public land and this sounds like your problem too, we’d like to talk.

Edgegenix is a multi-industry operational-intelligence platform; fire and smoke detection is our lead module and Land Encroachment is the next. Each module runs as its own tenant with its own console and data.