Renewable Energy Firm
EdgeVision (under NDA)

The Problem

A renewable energy firm required a real-time, offline-capable visual inspection system to assess equipment repair quality across 56 discrete stages. The existing cloud-based image classification pipeline was slow, costly, and lacked an explainability layer, causing significant operational bottlenecks. The model needed to run entirely at the edge without a stable internet connection while producing explainable outputs for verification.

What We Built

Sonder designed and trained EdgeVision — a lightweight, edge-deployable computer vision system built on ResNet-18 with transfer learning, optimized specifically for the 56-stage equipment 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

Sonder developed a clean data pipeline to deduplicate images, normalize labels, and augment underrepresented classes to resolve 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. This validated the model's resilience to lighting variations, motion blur, and partial occlusions.

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

The Outcome

EdgeVision reduced inference time to under a second with a 94% classification accuracy. Operating completely offline at the edge, the system eliminated cloud API dependencies, significantly reducing operational and processing costs.

Sonder's CTO led the end-to-end model optimization, architecture selection, and edge deployment strategy, ensuring the system met strict latency and accuracy targets.