How Landvex Validates Observations
I've seen a contributor try to submit a photo of a pothole taken from inside a moving car at dusk. The image was a brown smear with a suggestion of asphalt. The GPS pinged a shopping mall parking lot three kilometers from the claimed location. The timestamp said 3:47 AM in a city where the sun sets at 10 PM in July.
This is not rare. This is Tuesday.
Every observation that hits our system has to fight its way through four layers of validation before it becomes a finding anyone can trust. Not because we don't trust people. Because we've been doing this long enough to know what happens when you don't verify.
Layer 1: Collection validation — catch the garbage at the door
The quiXzoom app doesn't politely accept whatever you throw at it. It demands specific angles. It checks GPS against the claimed asset location — geotag yourself 500 meters from the nearest road and the app flags you instantly. It validates timestamps for plausibility. It rejects blurry photos, incomplete submissions, and locations that smell wrong.
No AI here. No magic. Just hard rules, enforced without mercy. A blurry photo of "maybe a crack" is worthless. The app tells the contributor exactly what's wrong and sends them back to do it properly.
This layer catches the obvious stuff. The stuff that would poison everything downstream if we let it through.
Layer 2: AI assessment — the machine looks, then admits doubt
Clean observations go to computer vision models trained on verified datasets. Crack detection. Corrosion spotting. Vegetation overgrowth. Lighting function. Each model spits out a confidence score: how sure am I about what I'm seeing?
Here's the thing most AI marketing won't tell you: these models are fragile. A crack detector trained on Stockholm asphalt will hallucinate cracks on Lisbon cobblestone. A lighting model gets confused at dusk. We know this. We calibrate per category, per surface, per lighting condition. The system knows its own blindness.
High confidence? Move to consensus. Low confidence? Flag for human review or send someone back for another look. The AI doesn't reject — it grades. And the grade determines what happens next.
Layer 3: Consensus scoring — nobody gets trusted blindly
Multiple observations of the same asset? The Consensus Engine weighs them. Not majority voting — weighted aggregation that considers who took the photo, how good it is, whether it matches history, and whether it contradicts the official register.
A contributor with 200 verified observations carries more weight than someone who just signed up yesterday. High AI confidence beats low confidence. An observation that aligns with known history beats one that comes out of nowhere.
The output isn't "broken" or "fine." It's: "Asset condition: degraded, confidence 87%, based on 4 observations, 2 validators, 1 contradiction with register." You get evidence strength, not a verdict. Interpret it according to your own risk appetite.
Layer 4: Provenance and audit — show your work
Every finding carries a complete chain: which observations contributed, what each layer scored, how consensus was reached, what the final confidence calculation yielded. Immutable. Inspectable. If you question a finding, we can trace it back to the exact photo, the exact GPS coordinate, the exact model version.
An observation turns out to be wrong? We recalculate its impact on every downstream finding. A model improves? We re-score history. Transparency isn't a feature we bolted on. It's the foundation the whole thing rests on.
Measuring quality — because hope is not a strategy
We track precision, recall, and confidence calibration continuously. Per model. Per category. When a model starts crying wolf — more false positives, poorly calibrated scores — the system screams and we retrain or replace it.
Quality isn't assumed. It's monitored. Like a patient in intensive care.
Request a pilot →
Data quality isn't sexy. It's the plumbing. But when the plumbing fails, everything else floods.