Shahul Ummarali · AI Systems Architect
Reference architecture

One platform behind every app

Each application is a thin workflow on a shared platform. The platform owns the hard parts once: identity, PHI handling, retrieval, evaluation and observability.

Healthcare payer AI platform architectureRequests flow from channels through an edge layer (gateway, redaction, injection screening) into use-case workflows, which call knowledge sources, tools and models. Code checks citations and routes cases to a human review queue. A quality and operations layer logs and evaluates every step.CHANNELSEDGEORCHESTRATIONMODELSKNOWLEDGE & TOOLSQUALITY & OPERATIONSInbound transactionsX12 278 · 837 · FHIR PASMember & providerportalsCall-center desktopReviewer workbenchnurse / physician UIAPI gatewaySSO · rate limitsIdentifier redactionPHI removed pre-modelPrompt-injectionscreenUse-case workflowsPrior authorizationDenials & appealsBenefits Q&APayment integrityHCC evidenceCitation check + routing rulesdeterministic codeHuman review queuea clinician makes every decisionClaude Opus 5.5Amazon Bedrock or Claude APIstructured output · prompt cacheClassic ML · XGBoostFWA outliers · overturn scoringPolicy libraryversionedHybrid searchpgvectorFHIR serverHAPI · SyntheaClaims &eligibility APIsCalculatorscode mapsLabeled test setsper applicationEvals in CIblock on regressionTracinglatency · cost per caseAudit loginputs, outputs, decisionsModel risk reviewdrift · re-labeling123456789
A request arrives from a portal, the reviewer workbench, or as an X12 278 / FHIR PAS transaction.

Highlighted boxes are the controls that make AI output safe to act on in a health plan. Teal arrows carry requests, orange arrows carry results back, the dotted arrow is logging, and the numbers follow one request.

Request flow

One request, end to end

  1. A request arrives from a portal, the reviewer workbench, or as an X12 278 / FHIR PAS transaction.
  2. The gateway authenticates the caller (SSO) and applies rate limits.
  3. Identifiers are redacted and documents are screened for prompt injection before any model sees them.
  4. The use-case workflow pulls the policy, clinical records, claims data and tools it needs.
  5. The workflow calls Claude for reasoning over text, or a classic ML model for scoring.
  6. Model findings come back as structured output with verbatim quotes.
  7. Code verifies every quote against the source and applies the routing rules.
  8. The case lands in the human review queue; a nurse or physician makes the decision.
  9. Every step is written to the audit log and traced for quality, latency and cost.
Decisions

Trade-offs I made, and why

DecisionChoiceReasoning
Agent or workflowFixed workflowsPayer processes are well defined and audited. A fixed sequence is easier to test, explain to compliance and debug.
Who decidesCode routes, people decideThe model produces findings. Routing thresholds sit in code where they can be reviewed and versioned.
Trusting citationsVerbatim match checkA cheap deterministic check catches paraphrased or invented evidence before a reviewer sees it.
RAG vs. fine-tuningRetrievalMedical policies change every quarter. Retrieval picks up a new version the day it's published.
HostingAWS, Claude on BedrockMost large payers already run on AWS under a BAA, so no new vendor review is needed.