Responsible AI
What the models do. Estimate the physical progress of construction work items from photographs, and flag likely workmanship defects and specification deviations.
What they never do. Identify people. Infer any human attribute. Produce a decision that affects a person without human review.
Known limitations. Accuracy degrades with poor lighting, motion blur, partial views, unusual materials, and viewpoints that differ substantially from previous days. Concealed work cannot be assessed once concealed. Percentages are estimates with error bars, not measurements. Defect flags are candidates requiring physical verification.
Human in the loop. Every output is advisory. Payment, acceptance, rejection, and dispute decisions are made by people. This is enforced in the product: no OneEye output triggers an automated financial or contractual consequence.
Fairness. Because the system assesses surfaces and assemblies rather than people, and because faces are removed before inference, there is no person-level inference to be biased about. We monitor for performance differences across regions, material types, and lighting conditions, and publish material findings.
Contest an output. Any customer or any worker at a documented site may contest a record at appeals@oneeye.in. We re-review with a human, correct the record where warranted, and log the outcome.
Oversight. Model changes affecting output are reviewed before release and recorded in /changelog.