HealthTech ISV Transforms Anthropic API Prototype into a Tenant-Isolated, HIPAA-Aware Clinical AI Platform with Claude on Amazon Bedrock
At a glance
A HealthTech ISV serving specialty-care clinics hosts more than 350 tenants, with identity, networking, and deployment on AWS. Its founder-built clinical AI tool was the exception: model requests went out to Anthropic’s own API, outside AWS and its controls, and some data routes exposed one clinic’s records to another. Shipping it meant risking a HIPAA breach.
New Math Data moved the workload from the Anthropic API to Claude on Amazon Bedrock, so model requests stay inside AWS and inherit the IAM roles and pipeline the company already runs, plus new PII guardrails and tenant isolation. One governed path replaced three ways of calling the model.
The company now has a platform ready for its planned pilot with 35 to 45 providers running 8 to 10 evaluations a day, and a model layer it governs from one place as it projects its AI user base will double within 12 to 18 months.
Industry
Use Case
Compliance & Governance, Responsible AI, Infrastructure as Code (Terraform), AI Platform Migration & Implementation
Solution implemented
- One Bedrock provider layer routes every Claude call, replacing three separate Anthropic API call patterns.
- Bedrock Guardrails mask PII on inputs and outputs and block prompt attacks and sensitive topic disclosure.
- IAM roles on Terraform-provisioned ECS Fargate tasks authorize Bedrock access, with no model API keys in application code.
- CloudFront and an internal Cognito-authenticated load balancer front Next.js on ECS Fargate, inside the existing network and pipelines.
- Tenant claims from Cognito tokens enforce isolation on every request, with all DynamoDB access moved server-side.
The value equation
- Clinical AI inference moves from the Anthropic API to Claude on Bedrock inside the company's own AWS estate.
- Model requests stay inside the company's AWS under its existing IAM and pipeline controls, plus new PII guardrails.
- Cross-clinic data exposure risk closed, with tenant isolation demonstrated against a simulated rogue actor.
- One governed model path replaces three inconsistent call patterns, with guardrail policies changed through configuration.
- Planned pilot reaches 280 to 450 clinical evaluations per day across 35 to 45 providers.
Company Snapshot
A HealthTech ISV provides practice management software to specialty-care clinics and is adding AI clinical review, with Claude moved from the Anthropic API to Amazon Bedrock.
Location
United States
Customer Situation
A HealthTech ISV builds practice management software for specialty-care clinics and hosts more than 350 tenants, with identity, networking, and deployment already on AWS. Its founder built an AI tool that reviews clinical evaluations, checks billing codes against payer rules, and drafts denial appeals. It ran as a sandbox prototype whose model requests went out to Anthropic’s own API.
That path was not fit to carry patient data. Engineering leadership wanted every model request inside AWS, under controls the company already operates, with guardrails changed through configuration. Some routes also lacked tenant isolation, a critical fix before release.
Across healthcare, AI tools that cannot show where patient data goes stay in the lab, while initial claim denials run at 11.8 percent of submissions (Experian, 2025).
NMD Solution
NMD reviewed and found that model inference was the one part running outside AWS: Claude was called three ways, one straight to the Anthropic API with an API key. The fix moved inference to Claude on Amazon Bedrock behind one provider layer, so it inherits the company’s existing IAM, network, and Bitbucket Pipelines, with Bedrock Guardrails for PII and prompt attacks, Amazon Cognito tenant claims, and Amazon CloudFront in front, all provisioned with Terraform.
What We Delivered
The company’s clinical evaluation review platform is deployed in its AWS development environment with every Claude request running through Amazon Bedrock, so model traffic stays inside AWS and the Anthropic API is out of the path. It inherits the company’s IAM, network, and Bitbucket Pipelines, adds PII guardrails, and keeps requests tenant-scoped, and a rogue-actor test shows one tenant cannot reach another’s data. The company holds the runbook, code, and Terraform and owns the production cutover. The build supports its pilot with 35 to 45 providers, and the company projects its AI user base will double within 12 to 18 months.
Ready to Transform Your HealthTech Business?
See how New Math Data can transform your HealthTech business with AWS-powered innovation.