omni:us: A cloud AI pipeline brings order to insurance claim processing

omni:us has built an AI pipeline on Google Cloud that sorts insurance claims paperwork and automates processing. The Berlin-based firm is using Google Cloud to extend the robotic book scanning technologies of its parent company, Qidenus Technologies, to deliver value to insurance providers and ultimately, policy holders.

For providers, auto insurance claims processing is a labor-intensive and time-consuming process. Each claim unleashes a flood of paperwork. It falls to insurance underwriters to gather, validate, and process vehicle registration details, police reports, accident forms, photos, repair estimates, invoices, and other documents — all from various sources and in a variety of formats.

It is also a significant opportunity for omni:us. "This year alone insurance companies will spend about US$250 billion to just handle claims, not settle them," says omni:us co-founder and CEO of Sofie Quidenus-Wahlforss. Competition in auto insurance is fierce. Surveys credibly suggest that claims settlement performance is a key indicator of customer loyalty.

Using AI to streamline claims processing delivers two attractive benefits: shrinking the time required to settle a claim while reducing the costs of doing so. To tackle the problem, omni:us is integrating a hybrid back office of on-premises and cloud-based visual neural networks, natural language processing (NLP), optical character, and handwritten text recognition with the goal of gaining control over — and intelligence from — a vast amount of unstructured information.

Put another way, omni:us seeks nothing less than automating claims processing to help insurers settle claims within minutes instead of weeks by cutting in half the amount of time it takes to process claims.

The omni:us AI solution is already making headway within the industry. Operating in six European companies and the United States, the firm is conducting pilot projects with nearly 30 clients.

omni:us relies on Google Cloud to host and support much of its AI processing. "What I personally like is that Google Cloud opens up its engineering team for questions," says VP of Engineering and Data Science Nischal Padmanabha. "Our team has participated in Google master class programs in Berlin and elsewhere." As omni:us implemented its solution, Nischal recalls, "It's been quite seamless for us as a startup to use Google Cloud."

Diverse AI stack sorts unstructured data

To the data scientists at omni:us, insurance paperwork is so much unstructured data — pieces of a puzzle for which solutions are now at hand. The challenge is converting the forms, photos, receipts, invoices, and other materials into structured data. Converting unstructured data to structured data opens the door to standardizing how data can be semantically evaluated. To achieve this, the omni:us back office brings together a variety of discrete open source and custom-built AI microservices.

The process starts when omni:us customers submit documents via an API. A well-defined API schema provides clues to document structure and content. Optionally omni:us provides a manual UI. In either case, an omni:us classification model determines the nature of the documents.

Depending on the data source — whether digital or scanned — omni:us aligns forms into global templates using computer vision processing, including Convolutional Neural Networks (CNNs). Modeled after the mammalian visual cortex, CNNs consist of a set of filters that are trained with data to identify patterns and classify documents as they are repeatedly scanned. With each scanning pass, the network detects and learns increasingly useful detail that culminates in a composite layer that is classifiable as a discrete form.

Next comes semantics information extraction. omni:us is developing a suite of NLP and computer vision AI models to understand text along with their positioning and relevance in terms of context. This involves supporting printed and scanned documents as well as handwritten forms via a state-of-the-art handwriting recognition model designed and built in-house.

The AI nets deliver results that are then validated by insurance claims professionals for settlement or further investigation, including fraud checking. "It is noteworthy that, for one of the handwritten claims cases for a car, the model achieved a handwritten character error rate of 7.25 percent," says Nischal.

Buckets reduce deployment time

omni:us has a hybrid platform: it captures data on-premises and processes it using cloud AI. To create its multi-layer, large scale neural networks, the omni:us data science team uses open source TensorFlow and PyTorch.

For its European clients, the firm chose Google Kubernetes Engine (GKE) to orchestrate the services and Container Registry in Google Cloud to store its containers. For its U.S. clients who have legacy AWS solutions, omni:us rolls out new features first on Google Cloud and then migrates them. omni:us also supports on-premises deployment for insurance customers by helping customers with installation of GKE on on-premises infrastructure.

Its loosely coupled AI microservices are packaged in buckets as container images. "We use buckets often," says Nischal. "Buckets help us mount the models. We can change the buckets in real time when we want to change the model, all without redeploying the cluster."

The tools have helped the team realize impressive gains in engineering productivity, reducing the time it takes to deploy an AI pipeline from two days to four hours.

omni:us relies on Compute Engine to run its training and other operational services. The firm's training models, backup configurations, and other data it elects to persist are relegated to Cloud Storage. Cloud Load Balancing serves as the entry point to the omni:us Kubernetes cluster to sustain high network availability and distribute resources under network load. To preempt DDoS attacks, omni:us configures the security layer on the load balancer to discern requests by TCP, SSL, or HTTPS.

Cloud Projects and Cloud Identity to protect client and personal data in compliance with the General Data Protection Regulation (GDPR)." data-pin-nopin="true">

Cloud privacy tools

omni:us uses a variety of Google Cloud solutions to support data privacy. "Insurance data is very private and GDPR compliance is strict. Meeting both requirements is a challenge because any data breach would essentially mean a big blow and substantial fines for the organization," says Nischal. "All of our services in the Kubernetes landscape run in zones that are GDPR compliant."

omni:us implements its own VPC with custom subnets in Google Cloud, configuring firewall rules to restrict communication to only those channels required by subnet instances. Cloud Projects help omni:us sequester customer data. "Cloud Projects help us manage the workloads of multiple customers without any chance of spilling over, which is critical in the insurance industry," says Nischal.

With Cloud Identity, omni:us created company-managed Google accounts that provide access from a variety of devices to cloud and on-premises services. Further protections are provided by Cloud IAM, which assigns and enforces more secure, role-based access to data and services.

On demand GPU processing, emerging standardization

Based on client preferences and location, omni:us works with a variety of cloud vendors. "Google Cloud is the cloud vendor that we primarily work with. And then we test everything on the other cloud platforms for different customers as required," says Nischal.

omni:us chose the Cloud GPU offering over others for its quality, support, and on-demand processing. "Unconstrained access to GPUs is important for us as a startup because we run so many experiments," says Nischal. "The quality of GPUs is fantastic, including their availability across geographical zones to support our far-flung clients." The firm plans to evaluate Cloud TPU to further accelerate processing.

omni:us is encouraged by what it sees as an emerging standardization around new AI technologies in a traditional industry such as insurance. As the barriers to adoption fall, the omni:us team credits the appeal and flexibility of technologies such as Kubernetes, docker containers, and storage management. "We have customers using these technologies on-premises and in the cloud," says Nischal. "This demonstrates how you can build scalable AI businesses in the B2B space."

Contributors to this story

Sofie Quidenus-Wahlforss: omni:us CEO and Co-founder. Sofie was a co-founder of Qidenus Technologies (handwritten text recognition) and winner of several awards including the Woman Technology & Research Award, Leonardo Best Automatization Technology Award, and Best Young Entrepreneur Award.

Martin Micko: omni:us COO and Co-founder. Martin has held executive positions at IBM, Sony, and Dell EMC.

Harald Gölles: omni:us CTO and Co-founder. A co-founder of Qidenus Technologies, Harald was a consultant at Coupon Future, GmbH, and Head of Engineering at meinKauf, GmbH.

Eric Pfarl: omni:us CXO and Co-founder. Eric is a co-founder of Qidenus Technologies, RepGain, and YSeed, and is a Strategic Advisor for Yushan Ventures, Ltd.

Stephan Dorfmeister: omni:us CFO and Co-founder. A co-founder of Qidenus Technologies, Stephan is the Managing Director at Deep Nature Project, GmbH, and at Conviva, Ltd.

Nischal Padmanabha: omni:us VP Engineering and Data Science. Nischal was a co-founder and Lead Machine Learning Engineer at Unnati Data Labs and Senior Software Engineer at RedMart and Associate Software Developer at SAP.