Document AI · Amazon Bedrock · AgentCore · Governed human review
AI Claims Triage
A governed claims triage agent built with Amazon Bedrock and AgentCore for document extraction, policy-grounded assessment and accountable human review.
Architecture: Amazon Bedrock + Amazon Bedrock AgentCore
The challenge
Claims teams must process varied documents quickly while preserving policy consistency, evidence, privacy and human accountability. Low-confidence extraction or reasoning cannot be treated as a final decision.
My end-to-end ownership
I independently delivered the engagement from initial discovery through production readiness. I owned requirements, workflow and experience design, architecture, document intake, extraction contracts, retrieval-backed checks, scoring, human review, frontend, backend, cloud infrastructure, security, testing, evaluation, observability, deployment and operational handover.
- Discovery and requirements
- Product and solution design
- Architecture and engineering
- AI and data workflows
- Cloud, security and CI/CD
- Testing, observability and handover
System design
Documents are securely ingested, classified and converted into schema-validated data. Retrieval-backed policy checks and deterministic rules produce evidence-linked triage recommendations, while confidence thresholds route exceptions to reviewers.
AWS claims architecture
The model assists classification, extraction and explanation, but policy, confidence thresholds and the final sensitive decision remain outside the model.
Encrypted uploads create immutable processing events with tenant, claim and document metadata.
Classifies documents and produces schema-aligned structured output from unstructured claims evidence.
The agent assembles the claim view, identifies missing evidence and prepares a source-linked triage recommendation.
Retrieves policy and procedure content with access filtering, citations and agentic retrieval for multi-step questions.
Restricts claims tools by user, role, purpose and action risk; approval is mandatory for sensitive outcomes.
Applies PII and grounding controls and measures extraction quality, policy adherence, escalation and reviewer agreement.
KMS encryption, private networking, IAM, Secrets Manager, CloudTrail, Step Functions, SQS dead-letter queues and an auditable reviewer console provide the surrounding production controls.
Production controls
Mandatory review for sensitive outcomes, structured outputs, source evidence, versioned model configuration, complete audit trails, PII controls, evaluation datasets and regression testing support safe adoption.