Metrics at a glance: auto-resolved, escalated, denials to human, cost per decision.
The challenge
Prior-authorization reviews are slow and repetitive, but they decide whether patients get care, so an AI can't just approve or deny on its own. The system has to be fast on clear cases and careful on the rest, with a full audit trail.
My role
I designed and built the agent, deployed it, and then ran a full security review and hardening of it.
What I built
• A LangGraph agent: retrieve the clinical criteria → a proposer suggests a decision → a critic challenges it → a judge decides
• Every denial and low-confidence case goes to a human reviewer with the reasoning attached
• Langfuse tracing on every step; deployed on Google Cloud Run; tools exposed as an MCP server
• Security review against the OWASP Top 10, then fixes: login, reviewer-only decisions enforced by Postgres row-level security, Google Secret Manager, audit logging and tested backups
The result
• 55% of cases resolved automatically at about $0.005 per decision
• 100% of denials reviewed by a human
• 25 security findings fixed, including two critical access-control flaws; GPT-4o resisted all 8 prompt-injection attacks
1 like · 0 views · Oct 27, 2026
LangGraph agent that resolves 55% of prior-auth cases at about $0.005 each, sends every denial to a human, and was hardened after an OWASP review.
