The problem
The situation this product exists to change, in the buyer's language.
Reviewing thousands of drone photographs per transmission line to find structural defects manually is slow and depends on a reviewer's attention holding up across a long, repetitive set of images.
Turning raw drone photography into a structured, per-tower inspection record — checklist, annotated evidence, and report — is a manual and time-consuming step.
Photos need to be sorted by tower before they are useful for inspection, which is its own manual task when a survey covers many towers.
A team piloting AI-assisted inspection needs a system that is honest about where it is, and is not yet, reliable enough to use unreviewed.
Approach
The direction the product takes to close that gap.
A YOLO-based detector (yolo26s-v9final, trained for 100 epochs) screens drone photographs for 13 classes, split between defects — broken insulator, broken conductor, corrosion, missing component, missing anti-climb device, bird nest, vegetation encroachment — and healthy components located in frame — insulator, conductor, arcing horn, vibration damper, corona ring, tower.
Photos are grouped into individual towers automatically, using EXIF GPS clustering and filename patterns, removing the manual sorting step before evidence can be reviewed.
Detections are annotated with bounding boxes and mapped onto a 46-point inspection checklist per tower, then compiled into a per-tower PDF report.
An operator dashboard supports uploading drone photography, reviewing the detected evidence, and downloading the generated report.
Any checklist point the system cannot assess from the available imagery is reported as NOT_ASSESSABLE rather than marked as passing. An earlier version inferred an 'OK' result from the absence of a detection; that was identified as a safety problem and removed. Absence of a detection is never reported as absence of a defect.
How it works
The path from a first action to the outcome it produces.
Photographs from a drone survey are uploaded through the operator dashboard.
The YOLO-based detector screens each image for the 13 defect and component classes it is trained on.
Images are clustered into individual towers automatically using EXIF GPS data and filename patterns.
Detections are annotated with bounding boxes and mapped onto a 46-point inspection checklist per tower; any point the system cannot assess is marked NOT_ASSESSABLE rather than passed.
A per-tower PDF report is generated for a qualified inspector to review before any operational decision is made.
Features
Grouped by the part of the workflow each set of features supports.
Fit
Who this is built for, what it changes, and what it runs on.
Questions
A first conversation can clarify fit, access, and what a quotation needs to cover.
On an independently audited benchmark of 30 licensed images from 10 sources, with no training-data overlap found, defect precision was 100% and defect recall was 50%; component precision was 92.9% and component recall was 20.3%. In practice, what the system flags is reliable, but it misses a substantial share of defects and components on imagery it has not effectively seen before, which is why it failed its own internal release gate on recall.
No. It is a pilot-stage screening tool. Every output requires review by a qualified inspector, and the system is not suitable for autonomous inspection, certified structural or electrical inspection, or a production-accuracy commitment.
It is reported as NOT_ASSESSABLE rather than as a pass. An earlier version inferred 'no defect' from the absence of a detection; that behavior was identified as unsafe and removed. Absence of a detection is never reported as absence of a defect.
Measured throughput was 30 images processed end-to-end — detection, tower grouping, annotation, checklist, and report generation — in 30.2 seconds.
The pilot covers detection across 13 classes, automatic grouping of images into towers, annotated evidence, a 46-point per-tower checklist, and a per-tower PDF report, supported by an operator dashboard for upload, review, and download. It does not automate coverage of all 46 checklist points, and it is scoped to a limited, human-reviewed proof-of-value engagement rather than production deployment.
The current path is a bounded, human-reviewed pilot. Request a quotation to discuss scope, imagery requirements, and what a proof-of-value engagement with your team would look like.
Get a quotation
Share the workflow, users, and constraints you're working with. We'll follow up with a scoped quotation.