Computer vision-based installation inspection for solar projects

Measure first, promise second

The initial request was for the system to count installed material from photos. A validation test across several projects and items showed this is not physically reliable: fasteners are hidden after installation, and counts from oblique angles are inaccurate. So we redesigned the system around the value that actually works: a photo quality gate, data integrity checks and proof of work-phase completion.

Technical implementation and business impact

Image recognition pipeline and on-site feedback

The system runs the installer's on-site photo through a quality check at upload. If the image cannot be assessed, it immediately asks for a retake, and after repeated failures it escalates to the project manager for review. A durable on-site reference board with a QR code provides team identification, a timestamp and a scale reference.

Three-state decisions instead of false certainty

The AI does not force a yes-or-no answer. For electrical connection checks, the output is “compliant”, “possibly non-compliant” or “not verifiable from the image”. When uncertain, it flags the case and routes it to a person instead of making up a result.

Measurable business impact and ROI

The system compares issued material with documented installed material using enterprise management data, making material loss visible at project and team level. A daily processing report and a management dashboard provide the overview. Photo-based proof of work-phase completion also serves as evidence in warranty disputes.

Data quality and integration

The audit also uncovered a critical data quality issue: the upload chain stripped photo metadata (timestamp, GPS), so fixing the source became a prerequisite for development. The solution connects to the client's own partner and work management system only through backend image recognition endpoints, without replacing existing interfaces.