How Landvex Validates Observations
Data quality is the foundation of everything Landvex does. A platform that delivers decision intelligence cannot compromise on the reliability of its inputs. Every observation that enters our system passes through multiple validation layers, each designed to catch a different category of error. The result is not perfect data — no system achieves that — but data whose quality is known, measured, and transparent.
Layer 1: Collection validation
Quality control begins at the moment of collection. The quiXzoom mobile app enforces structured data capture: contributors must photograph specific angles, confirm GPS location, and answer standardized questions. The app validates inputs in real time.
Photos are checked for focus, lighting, and framing. GPS coordinates are validated against the claimed asset location — an observation geotagged 500 meters from the nearest road is flagged. Timestamps are checked for plausibility. Observations that fail any check are rejected immediately, with guidance to the contributor on how to correct the issue.
This layer catches the most common data quality problems: blurry photos, incorrect locations, and incomplete submissions. It does not require AI or complex analysis. It requires clear protocols and disciplined enforcement.
Layer 2: AI assessment
Observations that pass collection validation proceed to AI assessment. Computer vision models analyze photos for the conditions of interest: cracks, potholes, corrosion, vegetation, lighting function, and more. Each model outputs a confidence score indicating how certain it is about what it sees.
The models are trained on verified datasets and calibrated per category. A crack detection model performs differently on asphalt than on concrete. A lighting model performs differently in daylight than at dusk. The system knows these limitations and adjusts confidence accordingly.
High-confidence AI assessments proceed to consensus scoring. Low-confidence assessments are flagged for additional review or additional observation. The AI does not reject observations — it grades them, and the grading determines what happens next.
Layer 3: Consensus scoring
When multiple observations of the same asset exist, the Consensus Engine combines them into a single scored finding. This is not majority voting. It is weighted aggregation that considers contributor reliability, observation quality, historical consistency, and official record alignment.
A contributor with a track record of accurate observations carries more weight than a newcomer. An observation with high AI confidence carries more weight than one with low confidence. An observation that aligns with the asset's known history carries more weight than one that contradicts it.
The output is a confidence-scored finding: "Asset condition: degraded, confidence 87%, based on 4 observations, 2 validators, 1 contradiction with register." This is not a binary judgment. It is a statement of evidence strength that decision-makers can interpret according to their risk tolerance.
Layer 4: Provenance and audit
Every finding retains a complete provenance chain: which observations contributed, what each validation layer scored, how consensus was reached, and what the final confidence calculation yielded. This chain is immutable and inspectable.
If a customer questions a finding, the full chain can be examined. If an observation is later found to be erroneous, its impact on downstream findings can be recalculated. If validation models are improved, historical observations can be re-scored.
Transparency is not a feature we added. It is a design principle that shapes every layer of the system.
Measuring quality
We measure validation performance continuously. Precision (how often flagged issues are real), recall (how many real issues are caught), and confidence calibration (whether stated confidence matches actual accuracy) are tracked per model and per category.
When performance drifts — when a model starts producing more false positives, or when confidence scores become poorly calibrated — the system alerts and the model is retrained or replaced. Quality is not assumed. It is monitored.
Request a pilot →