Relay Network ships secure genAI capabilities for regulated clients with AI Configs

Relay Network serves over 100 enterprise customers across healthcare, financial services, and specialty pharmaceutical sectors through a unique personalized business feed platform. The company enables client engagement through secure communication channels and data collection, serving as both a SaaS provider and a white-glove consultant to organizations that need to reach their customers with timely, relevant content.

Relay Network customers frequently faced a significant onboarding friction point: content creation. Their enterprise clients already had valuable content in various formats. However, transforming this long-form content into short, mobile-optimized messages suitable for Relay's platform required skills and training beyond the typical onboarding process.

"One of the first hurdles that either our clients or even our own implementation people have is getting a new client set up from square one," explained Chester O'Neill, Team Lead at Relay Network. "Rather than having to come in and key that in or send it to us and have us do it, letting them be able to just upload it and generate the content automatically became a critical need.”

The challenge was compounded by strict regulatory requirements. Healthcare and financial services clients often aren’t allowed to use consumer AI tools like ChatGPT due to concerns about compliance with PHI and PII regulations.

To make the feature valuable for customers, Relay Network also wanted to encode its best practices directly into the content generation process.

To build Content Pro, Relay Network knew they needed a solution that could support rapid iteration and avoid relying too heavily on engineering or developer resources for those changes. They also wanted to empower their product team to participate directly in AI development rather than funneling every change through engineering sprints.

The team also believed that lengthy deployment cycles would drain the energy from the creative process. "The magic is kind of lost in that whole iteration process. And then by the time it's actually deployed, they don't even really remember what they asked to change at that point," O'Neill noted.

Relay Network implemented LaunchDarkly AI Configs integrated with AWS Bedrock. AI Configs hosts all critical parameters, including model selection, temperature settings, guardrail IDs, trace levels, token limits, and prompts. These are then fed into Bedrock for output generation, with Amazon Nova Pro as the primary model.

Relay operates under strict HIPAA and HITRUST requirements, and by combining LaunchDarkly governance and versioning controls with AWS Bedrock’s built-in guardrails, the team helped to ensure that any PHI or PII accidentally included in uploads would be automatically caught and contained.

It took just a day and a half for the AI Configs implementation to get the basic AI functionality working. Within three weeks, the team had a functional and secure solution in the hands of beta clients. The entire project, from start to beta launch, took approximately two months, with most of that time spent on developing file upload and pre-processing capabilities.

Approximately seven team members, including one PM, an engineering manager, and four developers, actively collaborated on prompt development during the project. With engineering and development bottlenecks out of the way, the team reduced prompt iteration time from full-day deployments to seconds.

The team established and consistently met a 5-second generation benchmark while maintaining low error rates through schema validation and observability features from LaunchDarkly. By implementing AWS guardrails and leveraging their HIPAA certification and BAAs, Relay Network delivered a ChatGPT-like experience for regulated industries.

The success of Content Pro has increased Relay Network's confidence in AI development. The next phases of the project will focus on expanding Content Pro beyond basic summarization to support more dynamic, conversational interactions. The team also plans to experiment with industry-specific prompts and models.