By Michael Walker, CEO, Horus Technology
Our customer is a vital-records processing provider.
Their technology underpins vital statistics systems for multiple state health departments, powering services such as:
- Birth and death certificate processing
- Identity verification for public programs
- Epidemiological and public-health reporting
Every day, their platform handles large volumes of highly sensitive healthcare documents under strict HIPAA and state regulations.
To keep up with growing demand after the pandemic, they needed a new way to handle complex, multi-page medical and identity documents faster, more accurately, and at lower cost—without compromising compliance.
Working closely with AWS, we helped them build a Generative AI–powered document processing platform on Amazon SageMaker and a fully serverless data plane on AWS. The result is a platform that processes documents the same day, with a reviewer approving the output, and with far less manual keying than before.
What makes vital-records documents so hard to process?
The documents arrive as high volumes of PDFs scanned from different systems, resolutions, and layouts, and a single file can mention multiple patients and facilities. Facility names have many aliases across state and federal databases, and same-day deadlines and HIPAA requirements apply to all of it.
Our customer's quality control and identity document workflows sit at the heart of state-level vital records, so the pressure showed up in five places:
- High volumes of PDFs every day, often scanned from different systems, resolutions, and layouts
- Multiple patients and facilities mentioned in a single document
- Facility names with many different aliases across state and federal databases
- Same-day processing deadlines and HIPAA requirements
- Heavy, expensive manual adjudication whenever systems couldn't match data automatically
Existing approaches were hitting the wall:
- Traditional OCR and rules were brittle across layouts and states.
- Previous cloud-based pipelines for some workloads were slow, creating bottlenecks.
- Third-party OCR services and manual specialists were expensive to keep running.
Without change, they faced:
- Unsustainable processing costs
- Growing risk of SLA breaches and regulatory penalties
- Inability to onboard new states and programs quickly
- Competitive pressure from vendors adopting AI-driven automation
They needed a platform that could understand complex healthcare documents like a human reviewer, but run at cloud scale and speed.
Why did they choose Generative AI on AWS?
We chose Generative AI on AWS because it offered multimodal document understanding, a fully managed serverless data plane, native security aligned with HIPAA and state requirements, and rapid time-to-value through Infrastructure as Code. It let us focus on prompts, evaluation, and business rules while AWS handled undifferentiated heavy lifting like GPU provisioning, scaling, and multi-AZ resilience.
Together with our customer, we evaluated three main options:
- Enhancing the manual process with a more modern document management system
- Classic OCR + rules and regexes for pattern matching
- Custom-trained ML models running on self-managed GPU infrastructure
Each option had tradeoffs: high labor cost, poor robustness to layout changes, long MLOps build-out, or significant operational overhead.
We ultimately chose Generative AI on AWS because it offered:
- Multimodal document understanding with a vision-language model hosted on Amazon SageMaker
- A fully managed, serverless data plane built on Amazon S3, AWS Lambda, AWS Step Functions, Amazon DynamoDB, and Amazon OpenSearch Service
- Native security, encryption, and observability aligned with HIPAA and state requirements
- Rapid time-to-value through Infrastructure as Code with AWS CDK and automated CI/CD
How does the pipeline turn a PDF into structured data?
At the heart of the solution is a Generative AI pipeline on AWS that turns incoming PDFs into structured, searchable data ready for downstream systems. Documents are submitted to Amazon S3, orchestrated by AWS Step Functions, read by a vision-language model hosted on Amazon SageMaker, and stored in Amazon DynamoDB and Amazon OpenSearch Service.
1. Secure ingest on AWS
State systems and healthcare partners submit PDFs to Amazon S3, where they are encrypted with AWS Key Management Service (AWS KMS).
An Amazon SQS queue buffers work and decouples uploads from processing. A lightweight AWS Lambda function ("start processing") reads from the queue and kicks off an AWS Step Functions state machine.
AWS at work: S3 + SQS + Lambda + Step Functions provide a fully managed, event-driven backbone that absorbs bursts without manual capacity planning.
2. Orchestrated document pipeline with AWS Step Functions
The Step Functions workflow coordinates the full lifecycle:
- PDF → JPEG conversion – pages are split into images and normalized for AI processing.
- Patient and page analysis – map states parallelize processing across pages and patient segments.
- Text Extraction – each page or patient "chunk" is sent to the Generative AI model.
- Accuracy calculation & QC scoring – extracted fields are validated, scored, and prepared for human review where needed.
Every step emits structured logs to Amazon CloudWatch, making it easy to trace a single document from upload to final result.
3. Multimodal Generative AI with Amazon SageMaker
For the AI engine, our customer selected a leading vision-language model hosted on Amazon SageMaker.
The model ingests both the image of the page and extracted text, then outputs structured JSON with patient details, facilities, requesters, and key medical fields.
Asynchronous inference uses S3 input/output locations to handle heavy workloads without blocking callers.
Model metrics like latency and error rates are monitored via CloudWatch and power autoscaling policies.
This gives them the accuracy of a multimodal LLM with the reliability of a fully managed service.
4. Durable data foundation with DynamoDB and OpenSearch
Once extraction and validation are complete:
- Amazon DynamoDB stores file metadata, processing status, and the key patient fields needed for downstream systems and reporting.
- Amazon OpenSearch Service indexes the full text and selected fields, enabling fast search and reviewer workflows.
- Final JSON outputs are written back to an encrypted S3 output bucket for audit, reprocessing, or data-lake applications.
All storage is encrypted with AWS KMS, and access is controlled via least-privilege IAM roles.
5. Access, security, and governance
Security is non-negotiable in healthcare. The platform uses a defense-in-depth AWS design:
- Amazon Cognito enforces strong authentication and optional MFA for review applications and APIs.
- All APIs are fronted by Amazon API Gateway with HTTPS/TLS only, request throttling, and detailed access logging.
- VPC-isolated Lambda and SageMaker endpoints communicate via locked-down security groups and VPC endpoints; no database or search cluster is publicly accessible.
- AWS CloudTrail and AWS Config provide full audit trails and configuration visibility.
- Infrastructure is defined with AWS CDK, enforced with cdk-nag security checks, and deployed through a multi-stage CI/CD pipeline (dev → QA → prod).
Key AWS services at the core of the solution
Although we do not publish the full architecture diagram, the platform is built entirely on AWS and relies on the following managed services:
- Amazon SageMaker – Multimodal endpoint for Generative AI document understanding
- AWS Step Functions – orchestration of the PDF→JPEG→extraction→accuracy pipeline
- AWS Lambda – event-driven processing and micro-tasks
- Amazon S3 – input, async inference artifacts, and output archives
- Amazon SQS – decoupled ingestion and back-pressure handling
- Amazon DynamoDB – file metadata and extracted field store
- Amazon OpenSearch Service – search and analytics over extracted content
- Amazon API Gateway & Amazon Cognito – secure APIs and user authentication
- AWS KMS, AWS Secrets Manager, AWS IAM, AWS CloudTrail, AWS Config – encryption, secret management, and governance
- Amazon CloudWatch – dashboards, metrics, logs, and alarms for the entire workload
AWS isn't just the hosting environment—it's the engine that powers the entire GenAI application.
What changed with the new AWS-based platform?
Documents that were once keyed in by hand are now extracted automatically and a reviewer approves the result, so processing completes the same day. Automated extraction also replaces much of the manual keying, which lowers the manual effort and the third-party OCR and adjudication cost that came with it.
Turnaround time
- Before: documents were keyed in by hand, with queues and rework.
- After: documents are extracted automatically and a reviewer approves the result, so processing completes the same day.
This allows them to meet same-day SLA commitments even under heavy load.
Accuracy and quality
By combining multimodal Generative AI with a reviewer-in-the-loop workflow:
- Key patient, facility, and requester fields are extracted into structured data automatically.
- Extracted fields are validated and scored, and a reviewer approves the result.
- Reviewers spend their time on edge cases and new document types instead of routine documents.
Cost and scalability
Automated extraction replaces much of the manual keying, which lowers the manual effort and the third-party OCR and adjudication cost that came with it.
Queue-based ingestion and a serverless design let the platform absorb higher volumes as document counts grow.
Behind the scenes, CloudWatch dashboards and KPI standards track:
- API 5XX and latency
- Lambda p95 duration, errors, and throttles
- Step Functions failures and execution time
- SageMaker model latency and error rates
- DynamoDB and OpenSearch health
…ensuring the platform stays within its strict SLAs.
Transforming vital-records processing with AWS Generative AI
For our customer, this project is more than a technical upgrade—it's a strategic transformation of how vital records are processed:
- Citizens receive faster decisions and responses
- State health departments gain more timely, accurate data
- Compliance posture is stronger thanks to consistent encryption, access controls, and audit trails
- The platform is ready to onboard new states and document types without a complete redesign
Generative AI on AWS turned a labor-heavy cost center into an intelligent, scalable document platform that supports the mission of public health.
About Horus Technology and AWS
Horus Technology is an AWS-focused software engineering partner specializing in cloud-native, serverless, and Generative AI solutions for regulated industries such as healthcare and government.
We design and build production workloads on AWS using services like Amazon SageMaker, Amazon Bedrock, Amazon Textract, AWS Lambda, AWS Step Functions, Amazon DynamoDB, Amazon S3, and Amazon OpenSearch Service.
By combining deep AWS expertise with strong delivery practices, we help organizations modernize legacy workflows, automate document-heavy processes, and implement secure, audit-ready platforms that are ready for real-world production at scale.