How Our Claude Developers Built Clinical Documentation Automation for HealthOrbit
TL;DR: Claude developers for healthcare built HealthOrbit an AI-powered platform on Claude via Amazon Bedrock that lets physicians create, customise, validate, and publish clinical documentation templates without engineering involvement. Result: 85%+ minimum field-level exact match accuracy, support for 80% of common template use cases without manual intervention, and zero breaking changes to clinical documentation workflows.
HealthOrbit runs an ambient medical scribe platform, generating clinical documentation for physicians across specialties and geographies.
As the platform grew, HealthOrbit was receiving 5 to 10 template customisation requests daily from physicians, each requiring manual engineering review, prompt creation, validation, and deployment. That manual process became the bottleneck on how fast the platform could scale.
We partnered with HealthOrbit to put template creation directly in physicians' hands.
The Problem
Every new or adjusted clinical template went through the same manual queue.
- 5 to 10 template customisation requests arrived daily from physicians.
- Every request needed manual engineering review, prompt creation, and validation before it could deploy.
- Expansion across specialties and geographies multiplied the number of templates in flight.
- The manual review cycle delayed physicians waiting on templates and capped how fast HealthOrbit could scale.
HealthOrbit needed physicians to create and customise their own templates safely, without an engineer in the loop for every change.
The Solution
Our Claude developers built an AI-powered platform on Claude via Amazon Bedrock that lets physicians create, customise, validate, and publish clinical documentation templates independently.
Natural Language Template Creation
Physicians describe a template in plain language and Claude generates it automatically, with support for subsections, hierarchical layouts, and structured tables.
Document-Based Template Extraction
Physicians can upload an existing template as a PDF or image (JPEG/PNG) and the platform extracts a structured template from it automatically, using Amazon Textract for document parsing.
Automated Validation
Every generated template is checked against JSON schema compliance rules before it can publish, with dynamic header and footer generation carrying physician and organisation metadata.
Template Lifecycle Management
Backend APIs handle the full template lifecycle (create, read, update, delete), with version control, naming enforcement, and duplicate prevention built in.
The Architecture
AI and Document Processing
- Claude on Amazon Bedrock for template generation and validation
- Amazon Textract for PDF and image template extraction
Application Layer
- AWS API Gateway and AWS Lambda for the API surface
- Python REST APIs for template lifecycle management
- Amazon S3 for document and template storage
- DynamoDB for template metadata
Security and Monitoring
- Cloud-native security and monitoring services throughout
The Impact
- 85%+ minimum field-level exact match accuracy
- 85-90% extraction accuracy for digital documents
- Less than 10-15% hallucination rate
- 80% of common template use cases supported without manual intervention
- Zero breaking changes to clinical documentation workflows
Key Takeaways for Healthcare Platforms
Pitfalls to Avoid
- Keeping template customisation behind an engineering queue as the platform scales
- Accepting uploaded templates without automated schema validation
- Ignoring hallucination rate as a shippable metric for clinical documentation
What to Prioritise
- Put safe self-service in the hands of the people making the requests
- Validate structure automatically before a template can publish
- Design for zero breaking changes to existing clinical workflows
Build the Same Platform on Claude
If physician template requests are stuck in an engineering queue, our Claude developers can scope a similar self-service platform on Claude via Amazon Bedrock.