New Way Now: Captur delivers last-mile AI visual verification at the edge with Google Cloud

*Featured in this video:* Sumanas Sarma, CTO of Captur *Executive Summary:* Captur, which provides real-time visual AI for logistics and transportation, partnered with Google Cloud to deliver last-mile image verification, enabling users to validate photos on their devices, even offline. With Gemini, Captur developed a distillation architecture to train and deploy lightweight vision AI models that can analyze photos, provide feedback, and automate compliance instantly at the edge without the need for cloud-based processing or manual reviews. Google’s Gemini models also help test complex task logic and simulate data for real-world conditions to train and fine-tune model accuracy faster and more cost-effectively. Behind the scenes, Captur relies on Google Cloud to power its entire backend backbone with its lean engineering team, using Cloud Run, Pub/Sub, Cloud Storage, and BigQuery to run with maximum efficiency while minimizing operational overhead. Today, Captur handles more than 1 million sessions a day across thousands of mobile device models while building, training, and deploying its AI at scale. *Challenge:* Captur’s image AI needs to run on everything from the latest flagships to entry-level smartphones — without excessive data usage, battery drain, or performance lags. To deliver real-time verification at the edge, Captur wanted to find a way to optimize and manage model performance across thousands of unique devices while keeping operational overhead low for its lean engineering team. *Solution:* Partnering with Google Cloud, Captur used the Gemini platform to build a custom distillation architecture that helps train lightweight vision models that can run locally on mobile devices in milliseconds. Behind the scenes, Captur’s entire production stack runs on a fully managed Google Cloud backbone — including Cloud Run, Pub/Sub, Cloud Storage, and BigQuery — to deliver enterprise scale without the operational burden. Features like Cloud Run’s scale-to-zero compute also allow engineers to rapidly prototype new deployment setups or experiment without paying for idle capacity. In addition, Captur uses Gemini models to evaluate complex task logic and simulate data for real-world conditions to train models faster and more cost-effectively. Paired with BigQuery vector search capabilities, the team can now query visual data semantically and quickly pull training samples to fine-tune model performance and accuracy. *Results:* With Google Cloud powering how its models are built and trained, Captur now has a platform that can process and validate more than 1 million images daily — entirely on-device — in 30 milliseconds, even when offline. These capabilities are already delivering more reliable vision for operations in the wild, with one major delivery partner reporting a 30% reduction in “did not receive” claims, a 20% drop in unverified drop-offs, and 96% of their drivers correcting photos on the spot. *Key takeaways and highlights from our interview with Sumanas Sarma, CTO of Captur:* → “Before Captur, businesses either trusted blindly or waited for slow manual reviews. We create these real-time, mobile-friendly computer vision models. If a rider parks a scooter incorrectly or a driver takes a delivery photo that does not meet the rubric, we can catch that. We move the verification into the moment, on device.” → “Captur leverages Cloud Run for backend services. Scale to zero has been one of our most powerful benefits. It's allowed us to quickly experiment with options. BigQuery for data warehousing, Cloud Storage, Pub/Sub events, and then Gemini for model training. We also use Gemini for the multimodal AI capabilities. So, this integrated stack lets our very small engineering team deliver at scale.” → “The trick here is that we can't actually ship the large model on device. That's going to consume too much power. It's going to zap their network usage. So, we need to find a way to get the larger model's performance and get it running on the smaller model.” → “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. With the use of Terraform and Gemini, we were able to train our custom models. Google Cloud lets us do that without this army of Ops engineers that you’d normally need.” *Google Cloud products used:* The Gemini platform and models, Cloud Run, BigQuery, Cloud Storage, Pub/Sub *Learn more:* captur.ai