NovaScan: Real-Time Object Detection for Warehouse Logistics

The challenge

NovaScan Technologies operates 12 warehouse facilities across the UAE and Saudi Arabia, processing over 50,000 inventory items daily. Their existing workflow required workers to manually scan barcodes on each item — a process that created significant bottlenecks during peak hours and was prone to human error.

The core problem was threefold: barcode labels were often damaged or obscured, scanning required workers to handle each item individually (slowing throughput), and the existing system could not identify items without their barcode. NovaScan needed a solution that could visually identify products — by shape, color, packaging, and label text — without relying on barcodes.

They had spoken to several agencies and consultancies, but most proposed cloud-based solutions that would require constant internet connectivity — a non-starter in warehouse environments with unreliable WiFi. They needed on-device inference that worked offline, in real-time, and on standard iPads already deployed to their staff.

Our approach

We started with a two-week discovery phase. Our team visited two NovaScan warehouses, observed the existing scanning workflow, photographed over 3,000 unique product items under various lighting conditions, and interviewed warehouse staff about pain points. This fieldwork informed every technical decision that followed.

For the ML model, we chose YOLOv8 as the base architecture — optimized for speed without sacrificing accuracy. We trained it on our collected dataset of warehouse imagery (augmented to 15,000 images with rotations, lighting variations, and partial occlusions), then converted the model to CoreML format for on-device inference on iPad.

The app was built natively in Swift with SwiftUI. We chose native over cross-platform specifically for CoreML performance: the model runs on the Apple Neural Engine, processing camera frames at 30fps with an average inference time of 14ms. The camera pipeline uses AVFoundation for direct hardware access, with custom frame processing that skips duplicate frames to reduce battery consumption.

On the backend, we built a lightweight Supabase API that syncs inventory data when the iPad has connectivity. The app works fully offline — scanning, identification, and local inventory updates happen without any network connection. When WiFi becomes available, it syncs pending changes in the background.

Key technical decisions

Model size vs accuracy tradeoff: We tested three YOLOv8 variants (nano, small, medium). The small variant hit 96.8% mAP on our test set while keeping the CoreML model under 25MB — critical for fast app startup and low memory footprint on shared warehouse iPads.

Confidence thresholding: Rather than showing uncertain predictions, we implemented a three-tier confidence system. Above 90%: automatic match (green overlay). 70-90%: suggested match (yellow overlay, tap to confirm). Below 70%: manual lookup prompted. This reduced misidentifications to near-zero while maintaining speed for clear matches.

Incremental model updates: We built a model update pipeline that allows NovaScan to retrain the model monthly as new products are added to inventory. Updated models are distributed via MDM (Mobile Device Management) to all warehouse iPads without requiring app updates through the App Store.

Results and impact

The app launched as a pilot in two warehouses and delivered immediate results. Manual scanning time dropped by 73% — workers could now point the iPad at a shelf and identify multiple items simultaneously instead of scanning them one by one. Inventory accuracy improved from 94.1% to 99.2%, eliminating costly stock discrepancies.

After the pilot, NovaScan rolled out the app to all 12 locations within 60 days. The system now processes over 50,000 item identifications daily across their network. ROI was achieved in the first quarter — the reduction in scanning labor and inventory errors more than covered the development investment.

NovaScan has since contracted us for phase 2: adding shelf-level spatial mapping using LiDAR on iPad Pro, which will enable automated stock level detection and reorder alerts.

Lessons learned

Field research is non-negotiable for computer vision projects. Our warehouse visits revealed lighting conditions, item orientations, and edge cases that would have been impossible to anticipate from a desk. The two weeks we spent on discovery saved at least a month of rework.

Offline-first is the right default for enterprise apps. Even in environments with WiFi, network reliability varies. Building offline-first and adding sync later is far easier than retrofitting offline support into a cloud-dependent app.