Selected work

Stateful AI for healthcare claims and care operations

NYIC leadership experience engineering governed claims agents, typed services, durable execution, and sensitive-information controls.

Engineering lessons

How can several AI-assisted checks contribute to one accountable case workflow?

  • Give agents and services bounded responsibilities within a durable case state.
  • Separate concurrent checks, dependent actions, memory, and the authoritative case record.
  • Connect governed retrieval and tool permissions to production traces and release controls.

Context and professional contribution

Healthcare claims operations require coordination across records, policies, providers, and human reviewers. A case can contain incomplete evidence, conflicting findings, or work that must continue asynchronously after a downstream response arrives. An AI workflow therefore needs dependable transaction state and precise information access alongside its reasoning capability.

This study describes NYIC leadership experience building a production claims and care-operations platform on Azure. The documented role covered agent implementation, data preparation, tool services, access controls, asynchronous workflows, observability, and delivery infrastructure. The organization and individual case information remain anonymous.

The implementation connected specialized agents with governed analytical and operational services. It preserved a distinction between conversational context and authoritative claim state, helping the workflow remain inspectable when a case required review or recovery.

Assign bounded responsibilities to agents and services

Python agents built with Microsoft Agent Framework ran through Microsoft Foundry Agent Service. Separate agents supported claim review, provider context, explanations, and exception follow-up. Managed sessions held conversation history, and Azure Cosmos DB supported persistent agent memory.

Python and PySpark pipelines in Azure Databricks standardized claims and provider information, validated records, and prepared scoring features. Validated datasets were published to Snowflake on Azure for governed analytics and retrieval. Claim identifiers, reporting timestamps, and source lineage were preserved across these stages.

Model Context Protocol services handled claim scoring, rule checks, provider lookups, and explanations through typed input and output contracts. Business-critical calculations and rules remained in the services, while agents gathered context and coordinated execution. This boundary supported independent review of the operational rules used in a case.

Govern retrieval and tool authority

Azure API Management secured the MCP endpoints through Microsoft Entra ID token validation and scoped backend credentials. Permissions were assigned by agent and operation, separating read-only analysis from workflow updates. Approved agents and MCP services were registered in Azure API Center for discovery and version tracking.

The claims knowledge layer used Azure AI Search over policy, clinical, and operating documents. Searchable content and filterable metadata narrowed evidence by claim, provider, document type, and business context before it entered the agent workflow. This helped align retrieved material with the specific case under review.

The architecture illustrates why tool access needs to be defined independently from a model’s instructions. A proposed operation remains subject to service authorization and validation. The identity and scope attached to the request determine whether the workflow can perform it.

Coordinate concurrent checks and dependent work

Independent claim checks ran concurrently, while dependent steps followed sequential workflows. Ambiguous or conflicting findings passed through a supervisor agent before presentation to a reviewer. This gave the system an explicit path for reconciling findings without silently discarding disagreement.

Longer-running work used Azure Event Hubs through its Kafka-compatible endpoint, Service Bus queues, Event Grid, and Azure Functions orchestrated with Durable Functions. Bounded retries and idempotent writes supported recovery. Dead-letter queues and replay handlers allowed failed work to be resumed without duplicating claim decisions.

These execution controls addressed ordinary distributed-system failures. A worker can restart, a dependency can respond late, or a network connection can disappear after an update succeeds. Stable operation identities and inspectable recovery states help distinguish an unfinished action from a completed action whose response was lost.

Separate memory from the record of the case

Authoritative claim state was stored in Azure Cosmos DB, supporting relational information in Azure Database for PostgreSQL Flexible Server, and documents and artifacts in Azure Blob Storage. Conversational memory was isolated from transaction records through separate containers and access permissions.

This design prevented a conversational summary from becoming the sole record of an operational decision. A memory can help an agent resume context, while the transaction record establishes the actual case state. Maintaining that distinction is important when users correct information or a workflow must replay a failed step.

The application also inspected protected health information before model calls and before responses were returned or persisted. Azure AI Language PII detection was combined with healthcare-identifier rules. Least-privilege access used Entra ID and Azure role-based access control, with customer-managed encryption keys in Azure Key Vault for persisted claim information.

Observe and release the production system

OpenTelemetry instrumentation covered incoming APIs, retrieval, routing, MCP calls, model invocations, and final responses. Foundry traces were correlated with Azure Monitor Application Insights and Log Analytics. This connected failures across service boundaries and supported investigation of the full workflow.

Supporting FastAPI and MCP services ran on Azure Kubernetes Service. Terraform provisioned infrastructure, and GitHub Actions promoted releases through development, test, and production. Automated tests, approval gates, and rollback procedures were part of the delivery pipeline.

Evaluation of the workflow should examine rule correctness, evidence support, information access, duplicate prevention, and the quality of escalations to reviewers. These measures connect the platform’s technical behavior to the case decisions and operating responsibilities it supports.

Documented result and transferable lesson

The delivered system connected specialized claims agents, governed evidence, deterministic services, and durable asynchronous execution. Its controls addressed both sensitive-information handling and the integrity of operational state. Human review remained part of the path for ambiguous or conflicting findings.

The central lesson is that production healthcare AI requires coordinated control over evidence, authority, and execution. A useful explanation must remain connected to the correct case, and a correct recommendation must still pass through the authorized process before it changes that case’s operational state.