Agentic SCRA by Ippon
What if your SCRA compliance process ran itself?
Document review & AI extraction
Claude API and OCR pull Name, Branch, and Active Duty Start Date from military order PDFs -- regardless of layout variation. A human representative reviews and confirms every extracted field before the workflow advances. LLMs never act without human confirmation at this stage.
DMDC verification & duty date reconciliation
Structured API calls to the official Defense Manpower Data Center confirm service member identity. A traditional rules engine then reconciles dates between uploaded documents and DMDC records. No LLM involved -- these steps are faster, cheaper, and cleanly auditable.
Per-account AI evaluation with mandatory human gate
Amazon Bedrock evaluates SCRA eligibility rules against every linked account (credit cards, mortgages, auto loans) and generates a compliant proposal -- Enroll or Decline, with rationale. A human operator must review and approve each decision before fulfillment. The AI never acts unilaterally on a regulated account.
Forensic audit trail -- append-only by design
Every LLM invocation, model version, exact prompt hash, token payload, and retry attempt is written to an unalterable RDS Postgres table. Complete decision replay is available on demand. Exportable for OCC, FDIC, or DOJ requests. A kill switch lets an administrator instantly halt the AI engine -- all active cases pause via Step Functions with zero data loss.
Full 4-stage SCRA workflow validated end-to-end
Complete pipeline -- document extraction, DMDC verification, duty date reconciliation, linked account discovery, and per-account AI decisioning -- deployed on AWS ECS Fargate with Step Functions orchestration. PII tokenization and kill switch validated. Demo-ready for banking prospects.
Banking-grade security architecture completed
PII tokenization strips SSNs, account numbers, and names before any LLM call -- foundational models never see real customer data. Regex injection scanning, Bedrock Guardrails, rule registry restriction, and per-case token budgets all validated in staging.
DMDC API integration and Step Functions orchestration completed
Deterministic DMDC verification and duty date reconciliation stages operational. AWS Step Functions workflow handles human-in-the-loop pauses via task tokens -- active cases survive system restarts with zero data loss. Append-only audit ledger schema finalized.
