The challenge
Aiberry needed to build out a software platform to provide quantitative analysis, mental health insights, and risk scores to health care providers in real time during short interviews from captured media.
The system needed to be able to process imagery, audio, speech, and other media in real time and leverage Large Language Models (LLMs) for analysis and risk scoring. The system needed to be globally available, accessible, and redundant with low latency. Onboarding of new providers needed to be quick and seamless. Modifications and enhancements to the underlying AI/ML cloud infrastructure needed to be fast, simple, and reliable. Large Language Models (LLM) needed to be flexible and adapt to new media, data, and usage scenarios over time. Costs needed to be managed, observed, and scalable.
The solution
Sela’s solution relied heavily on machine learning (ML) services on AWS. Media streams were converted to raw transcripts, with ML inference used to predict patient suicidality using:
- Lambda + Step functions
- ECS
- SQS
- DynamoDB
- Sagemaker
- AWS Comprehend
A simple user interface was built for clinicians and patients and deployed to a secure, multi-account AWS environment using containers running on ECS. AWS SSO provided secure roles and permissions for users. Infrastructure builds across tenants and regions were automated with AWS CDK, Gitlab, and CodePipeline.
The results
Aiberry is now able to quickly adapt their system to changing requirements, new research, and new technology. Documentation and training on the simplified end-user application make it easier to user and understand. This translated into greater subscriber growth via faster onboarding and a better user experience with fewer support issues and performance problems.
The system performs faster and costs less per user to operate. The product is easier to innovate and deploy.
Accessibility
