A national insurance data ecosystem aggregates sensitive records from every insurance company in the country. Perimeter security doesn’t address insider threat or an AI-era access pattern: autonomous agents, not just people, now request access to that data. Binary role-based access control has no way to reason about that.

I was sole architect for the AI Trust Fabric: a Zero Trust framework (NIST 800-207) structured as PAP / PDP / PEP / PIP — policy authoring, real-time policy decisions, enforcement at every access point, and context aggregation (identity, device posture, location, time, data classification, session history). On top of that sits an Adaptive Trust Engine, computing a continuous trust score per request rather than a binary allow/deny, driving ALLOW / STEP-UP-AUTH / DENY decisions.

The Zero Trust decision pipelineA left-to-right pipeline of five stages: Identity (agent and human), Policy (policy administration and information points), Decision (policy decision point plus adaptive trust score), Enforcement (policy enforcement point), and Telemetry (audit trail and traces). A feedback loop runs from Telemetry back to Policy, labelled continuous evaluation, showing that trust is re-evaluated rather than granted once.Identityagent + humanPolicyPAP · PIPDecisionPDP+ trust scoreEnforcementPEPTelemetryaudit · tracescontinuous evaluation — trust is re-scored, not granted onceevery request · human or agent
Access decisions flow through policy, not around it — and telemetry feeds back into the next decision.

Above the access layer sits an AI governance layer operating on agent behaviour itself: per-agent rate limiting, action-scope enforcement, prompt injection detection, a full decision audit trail, and human escalation for high-risk actions. Zero Trust ensures no unauthorised access; the governance layer ensures authorised agents stay within the boundaries of their intent.

A RAG policy-intelligence layer (LangGraph orchestration, running on Ollama) lets analysts and agents query insurance regulations in natural language — fully offline, because certain environments have no internet egress. That was a hard constraint, not a preference: without offline inference, the AI governance layer couldn’t have shipped at all.

Every AI-assisted access decision carries a traceable rationale — which policy matched, which trust signals fired — and AI agents have their own workload identity, distinguishable from human actions in the audit trail.

This work is now a registered design with the Government of India. I’ve written up the architecture and what I learned building it in Building the Trust Fabric.

Stack: Entra ID, ZTNA, IAM, CrewAI, LangGraph, Ollama, MCP, Kubernetes.