Captur

Captur: Cutting delivery claims 30% and elevating edge AI compliance with Google Cloud

Results on Google Cloud
  • 30% fewer "did not receive" claims in the first week

  • 96% of drivers corrected delivery photos on the spot

  • 20% fewer invisible package deliveries recorded

  • 1M+ images processed daily across 7K phone models

Using Google Cloud to power real-time image AI at the edge, Captur is helping eliminate proof-of-delivery disputes and city compliance gaps.

Bringing real-time visual verification
directly to the edge

In the fast-moving logistics and transportation sectors, couriers and riders take millions of photos daily to verify successful package drop-offs or correct e-scooter parking. Local governments are also increasingly selective about which micromobility operators they invite onto their streets. Captur offers on-device visual AI for logistics and transportation, validating photos from users in milliseconds to help verify proof of delivery and parking compliance.

"Before Captur, businesses either trusted blindly or waited for slow manual reviews," says Sumanas Sarma, Captur’s CTO. For example, issues like a package delivery dropped in a bush or an e-scooter parked outside of designated bays were caught far too late, leading to costly disputes, customer refunds, or city compliance gaps.

When a company uses a photo as proof of something—a delivery, a parked bike, or a completed vehicle inspection—Captur analyzes camera frames in real time on mobile devices to provide instant feedback before an image is uploaded. “When the courier is taking the photograph, we can nudge them to say, ‘Is this even a drop off location?’” Sarma notes. “We move the verification into the moment, on device. That will allow both the system and the recipient to get a good photograph.”

However, securely deploying computer vision models locally presented an immense challenge for Captur’s lean engineering team. Every mobile device, from the latest flagship to an entry-level smartphone, is essentially a distinct platform with its own set of constraints and performance trade-offs. “The trick here is that we can’t actually ship the large model on device,” Sarma says. “That’s going to consume too much power. It’s going to zap their network usage.”

Building on Google Cloud gives Captur a complete AI stack: AI infrastructure, leading models, data management capabilities, and developer tools. This foundation allows the team to build, train, and distill powerful models at scale that run entirely on-device, with minimal operational overhead.

Before Captur, businesses either trusted blindly or waited for slow manual reviews. We move the verification into the moment. That will allow both the system and the recipient to get a good photograph.

Sumanas Sarma

CTO, Captur

Powering thousands of AI models on any device

By running inference directly on-device, Captur provides zero-latency feedback without costly or slow cloud-based processing or manual reviews—ensuring verification happens instantly at the moment of capture, online or off. To bridge the gap between large, resource-heavy AI models and the physical limitations of mobile hardware, Captur developed a custom distillation architecture using the Gemini platform.

“We need to find a way to get the larger model's performance and get it running on the smaller model. That's the purpose of distillation here,” Sarma emphasizes. “We'll have the smaller model learn from what the large model is doing. We have to pick and choose the different models that will be put together to create this one final model that will actually run on device.”

Behind the scenes, Captur relies on Google Cloud to run its backend microservices and AI pipelines with maximum engineering efficiency. Microservices are deployed to Cloud Run, where the scale-to-zero capability allows Captur to rapidly prototype new deployment setups or test experimental models without incurring charges for unused background compute. Asset storage resides in Cloud Storage, real-time events stream through Pub/Sub, and analytical data is centralized inside BigQuery.

"Google Cloud lets us do that without this army of ops engineers that you'd normally need," Sarma reflects. "Our production stack runs entirely on Cloud Run, Cloud Storage, BigQuery, and Pub/Sub. Now, we have over a million images a day, and we really had the platform support us as we scaled."

This approach has allowed Captur to combine edge-based model execution with a serverless backbone, transforming everyday mobile devices into high-precision verification hubs. In the first week of deployment with a major delivery partner, for instance, Captur drove a 30% reduction in "did not receive" claims, while unverified package drop-offs fell by 20%. Crucially, 96% of drivers corrected their photos on the spot when prompted, leading to improved photos for the intended recipient and better verification for the delivery partner.

Google Cloud lets us do that without this army of ops engineers that you'd normally need. Our production stack runs entirely on Cloud Run, Cloud Storage, BigQuery, and Pub/Sub. Now, we have 750,000 images a day, and we really had the platform support us as we scaled.

Sumanas Sarma

CTO, Captur

Keeping real-time image capture in sharp focus

To maintain peak accuracy across an expanding fleet of supported mobile devices, Captur also uses Google Cloud to continuously stress-test model performance and pull real-world edge cases into its retraining loop.

“We face classic ML ops challenges, deployment, monitoring, making sure the model doesn't drift over time,” Sarma notes. “All of this is with the context that the model is essentially a guest on someone else's phone.”

Before building or deploying new models to mobile hardware, Captur utilizes Google’s Gemini models to test task feasibility, prototype logic, and evaluate rubric complexity. “We use Gemini to test what's feasible before committing to on-device models. If Gemini is struggling, we know the task is framed in a complex way,” Sarma says. “Previously, we would have to iterate. This allows us to automate performance and make sure the model is robust when it gets shipped.”

Gemini’s multimodal capabilities enable Captur to generate synthetic dataset variations that reflect diverse weather and environmental conditions, such as snow, rain, or twilight. In addition, using Terraform and the Gemini platform to automate model training gives Captur a reproducible, scalable path for continuously optimizing its smaller edge models.

While synthetic data expands training coverage, capturing genuine on-the-ground exceptions remains essential for fine-tuning. Beyond serving as a central repository for millions of image files, according to Sarma, BigQuery’s vector search simplifies how Captur identifies visual anomalies.

"We've also made heavy use of BigQuery," he shares. "We can query semantically and that would allow us to pull in this data to train the model to improve." By executing semantic queries directly across their data warehouse, Captur can instantly isolate specific image capture issues—such as poor lighting or newly installed bike racks—and routes those specific assets back into the model pipeline.

Looking ahead, Captur sees its partnership with Google Cloud as a way to continue pushing the boundaries of edge intelligence. "The next leap I think is making the feedback more specific," Sarma reflects, "As these models become small enough and efficient enough, we can start combining this visual understanding and then a planning layer that decides what action to recommend. With Google Cloud, Captur is creating a new way to capture the world."

We use Gemini to test what's feasible before committing to on-device models. If Gemini is struggling, we know the task is framed in a complex way. Previously, we would have had to iterate. This allows us to automate performance and make sure the model is robust when it gets shipped.

Sumanas Sarma

CTO, Captur

Captur is AI infrastructure for enterprise mobile-apps that instantly verifies user-submitted photos for quality and accuracy, running complex machine learning models entirely on-device without cloud inference.

Industry: Startup

Location: United Kingdom

Products: Cloud Run, BigQuery, Cloud Storage, Pub/Sub, Gemini platform, Gemini models

  • Produits Google Cloud
  • Parcourez plus de 100 produits. Les nouveaux clients bénéficient de 300 $ de crédits gratuits pour exécuter, tester et déployer des charges de travail. Tous les clients peuvent utiliser plus de 25 produits gratuitement, dans les limites mensuelles spécifiées.
Google Cloud