Edge AI Proof of Delivery: Offline Computer Vision

Guarantee the integrity of proof of delivery across a nationwide last-mile operation handling about 1.000,000 parcels a day. Drivers record each delivery in the company's mobile app with a storefront photo and a recipient signature, but blurry, random or covered photos and signatures made of dots or scribbles were undermining trust, generating customer complaints and support rework. 

Goals: (1) automatically classify, in real time, whether the photo shows a valid building facade and whether the signature is genuine; (2) give the driver instant feedback so the error is fixed before the delivery is closed; (3) run fully offline on low-end devices, in areas with no connectivity; (4) reduce delivery-proof complaints and fraud; and (5) cut the support cost of handling inconsistencies.

Results. 431,000 invalid signatures (7.8% of records) and 353,000 invalid facade photos (23.4%) caught and corrected at the moment of delivery. Delivery-proof complaints down 15%; estimated savings of 350k in rework and complaint handling over 7 months; investment of about R$ 20,000. The solution runs at full national scale.

Next steps. Reuse the compression and monitoring approach for other offline AI use cases, such as real-time quality control at other logistics stages.

• Proprietary last-mile mobile app – delivery status, proof of delivery, geolocation, routing

• TensorFlow Lite – on-device, offline inference

• Image classification models – valid vs. invalid facade photos; genuine signature vs. scribble

• Model compression and quantization – from ~90 MB to under 3 MB

• Synthetic data generation scripts – scribbles, lighting and angle variations for training

• Data storage and monitoring structure – tracking classifications at ~1.000,000 parcels/day

• Agile delivery – Scrum and Kanban, beta releases with driver feedback

Three hard constraints had to be met at once. (1) Device reality: many drivers use low-capacity phones, so the model had to shrink from about 90 MB to under 3 MB through compression and quantization, without losing accuracy, and run fast on very different hardware and operating systems. (2) No connectivity: validation had to work fully offline, so every decision is made on the device. (3) Real-world variety: a valid facade can be a house, a condominium gate or a building, under very different light and regional architecture. Building the training set was manual and slow, so the operations team helped label real photos and scripts generated synthetic scribbles and adverse conditions. Rolling out to the full operation, about 500,000 parcels a day, also required a scalable structure to store and monitor classifications. Finally, drivers had to see the feedback as guidance, which required careful UX and operational training.

Context. Total Express is one of Brazil's largest logistics and e-commerce delivery companies, handling about 500,000 parcels a day nationwide. Its proprietary last-mile app gives drivers real-time delivery status, signature capture, recipient identification, geolocation checks, non-conformity reporting and daily routing.

Challenge. Following market benchmarks, the app added a facade photo as proof of delivery, which also protects recipients who prefer not to share personal data. But some drivers submitted blurry, random or covered photos and signatures made of dots or scribbles. This damaged the credibility of the proof, created friction with customers and increased rework.

Solution. We embedded AI image classification in the app to validate photos and signatures in real time. Two models classify facade photos (houses, condominiums and buildings, with regional and lighting variations) and signatures (genuine vs. scribble) as valid or invalid. When the result is invalid, the driver gets instant feedback and must redo the proof before closing the delivery.

Engineering. Requirements were strict: lightweight, fluid on devices with very different hardware and operating systems, and fully offline. The first model was about 90 MB; compression and quantization brought it under 3 MB with no loss of effectiveness, and it reached over 95% accuracy in controlled tests, improving progressively in the field. Training combined real samples, labeled collaboratively with the operations team, and synthetic samples generated by scripts.

Method. Pain-point mapping with customers and operations, multidisciplinary sessions to capture regional differences, Scrum and Kanban, driver-journey design, usability tests, gradual beta rollout, continuous review of positive and negative samples, and training for operations teams.

Brazil (nationwide last-mile operation)