UK Renewable Energy Firm
EdgeVision

The Problem

A UK-based renewable energy firm had deployed a visual inspection pipeline to assess turbine blade repair quality across 56 discrete repair stages. The existing system used GPT-4o for image classification — sending photos to the cloud, waiting for a response, and logging the result. On paper, it worked. In practice, it was creating serious operational bottlenecks.

Each full inspection cycle took approximately four hours. Field engineers were forced to wait on-site while cloud calls completed, connectivity was unreliable at turbine sites, and the GPT-4o pipeline was costing far more per inspection than the client had budgeted. Worse, the system had no interpretability layer — when it flagged a repair as failing, engineers had no way to understand why, which eroded trust in the system.

The client needed an inspection system that could run at the edge, work without a stable internet connection, complete assessments in seconds rather than hours, and produce explainable outputs that field engineers would actually trust.

What We Built

Sonder designed and trained EdgeVision — a lightweight, edge-deployable computer vision system built on ResNet-18 with transfer learning, optimised specifically for the 56-stage turbine blade repair classification task.

The core architecture choices were deliberate:

• ResNet-18 + transfer learning — rather than fine-tuning a massive foundation model, we used a compact, well-understood architecture pre-trained on ImageNet and fine-tuned on the client's labelled image dataset. This gave us strong classification performance with a model small enough to run on edge hardware
• Offline inference — the model is packaged as a self-contained inference binary that runs on a ruggedised tablet or laptop without any cloud dependency. Connectivity is only needed to sync inspection logs, not to run the model
• Explainable AI (XAI) via Grad-CAM — every classification comes with a visual heatmap overlay showing which regions of the image most influenced the decision. Engineers can see exactly what the model is "looking at", which dramatically increased on-site trust
• 56-class multi-stage output — one model, one inference pass, outputs a pass/fail/review status for each of the 56 repair stages relevant to the image, rather than requiring 56 separate calls
• Confidence thresholds with human-in-the-loop routing — predictions below a confidence threshold automatically route to a human reviewer queue rather than auto-classifying, keeping the human oversight meaningful without slowing down high-confidence cases

Training & Validation

The client's existing image dataset needed significant cleaning before it was useful for training. Sonder spent the first three weeks on data pipeline work: deduplicating images, normalising labels across their inconsistent annotation scheme, and augmenting underrepresented repair stage classes to reduce class imbalance.

We trained across multiple ResNet-18 checkpoints and evaluated on a held-out validation set of 10,000+ real-world inspection images captured in genuine field conditions — not a clean studio environment. This was critical: field images have lighting variation, motion blur, and partial occlusion that a model trained only on clean imagery would fail to handle.

Final validation accuracy: 94% across all 56 repair stages.

The Outcome

EdgeVision reduced inspection cycle time from 4 hours to under a second for model inference, with total on-site time — including log sync and any human review routing — dropping by over 85%. Field engineers adopted the system quickly once the Grad-CAM overlays gave them a reason to trust the outputs.

The client's operational cost per inspection dropped substantially compared to the GPT-4o pipeline. The system now runs as the primary inspection tool across their active turbine sites, with the human review queue handling the ~6% of cases where the model's confidence falls below threshold.

Sonder also delivered a full model card and data sheet for the system, documenting its performance boundaries, known failure modes, and recommended maintenance schedule — making it auditable and supportable beyond the initial engagement.