Claims (Pre-Auth, Coding and Adjudication)
AI agents across the full claim lifecycle — pre-authorization, hospital coding, and payer-side adjudication — each grounded in the chart, never guessing.
Claims covers three stages in one app. Pre-Authorization checks a request against the payer's policy criteria before care, with a word-for-word chart quote for every criterion and numbers such as HbA1c decided by code. Coding & Prep reads a completed hospital stay and prepares what belongs on the claim, priced with the real CMS MS-DRG grouper. Adjudication reviews a submitted claim the way a payer would — pre-authorization on file, bundling, frequency limits and duplicates checked deterministically, then medical necessity reviewed with a verified quote — reaching a paid, denied or pended determination with a real CARC reason code. A verification step in every stage rejects any quote that is not really in the chart, and nothing is ever submitted, billed or paid automatically.
Without it, each stage is its own manual re-read of the same chart — staff checking policy by hand before care, a coder catching what was documented but not coded after discharge, and an adjudicator applying coverage rules with no consistent, auditable evidence trail — with over 5 billion claims processed in the US every year.
This isn’t a proof of concept.
What you’re about to see is a fully containerized, production/enterprise-grade solution — the same build that deploys to Kubernetes on any of the three major cloud providers: AWS, Azure, or Google Cloud. There is no gap between this demo and what ships to production.
Built in, not bolted on
Agent Governance
Every agent call runs inside a governor with its own time and token budget and a kill-switch — a runaway or misbehaving step is stopped automatically, not left to run up cost or produce a bad answer, and every call leaves an audit-trail row.
PII / PHI Safe
User input is screened for personally identifiable and health information before it ever reaches a model, so sensitive data doesn't leak into a prompt, a log, or a third-party LLM call by accident.
Prompt-Injection Safe
A dedicated classifier checks every input for attempts to hijack the agent's instructions before it's acted on — the kind of "ignore your previous instructions" attack that a plain chatbot has no defense against.
SQL-Injection Safe
Where a solution talks to a database, every generated query is checked against a strict allow-list before it runs — no destructive statement (drop, delete, update, alter) can reach the database, however it's phrased.
Observability — SSE, Phoenix & Live Logging
Every step an agent takes streams live to the screen as it happens (no blank-screen wait for a final answer), and mirrors into a self-hosted Phoenix tracing dashboard — full request timelines, per-agent spans, and real token/cost usage, visible in real time, not reconstructed after the fact from a log file.
QoS — LLM-as-Judge
Before an answer ever reaches the user, a second, independent model call checks it for safety and can block it outright; a separate quality pass then scores completeness, correctness and relevance in the background — real automated review, not a cosmetic "checking quality..." status line.
Self-Improving (RSI via SKILLS.md)
Agents record what they learn from real runs into version-controlled SKILL.md files, which future runs read back — the system gets measurably better at its job over time instead of staying frozen at its original prompt.
Bring Your Own Model (BYOM)
Switch the underlying model with one setting — a fully local, on-prem model, GPT, Gemini, Claude, Amazon Bedrock, or Microsoft Foundry — with no code change and no vendor lock-in. Every option is a genuinely working, tested path, not a stub.
Durable Memory Across Sessions
Give the agent a user ID and it remembers your last few questions and answers — a follow-up like "which of those spent the most?" resolves correctly days later, without re-explaining context every time you come back.
Tenant Info Isolation
That remembered history is scoped strictly per user ID at the database level — one person's session data is never visible to, or blendable with, another's, even on the same solution.
Role-Based Access Control
Every account carries a role — user, admin or owner — and the server checks it before a solution opens, so a correct password alone isn't enough. Your username and role sit beside Log out, and every run starts with “Applying RBAC based access control” in the live stream.
What it’s built with
Python backend, served with Uvicorn
A FastAPI service running on Uvicorn handles every request — the same production Asynchronous Server Gateway Interface (ASGI) stack used across the whole platform, not a notebook or a prototype script.
FastMCP tool server — MCP v2 support
Database access, document tools and integrations are exposed through the Model Context Protocol via a dedicated FastMCP server running on MCP v2, so agents call real, typed tools instead of hand-rolled function stubs.
Large MCP Payload Support
We go beyond the default MCP payload of 5MB. We support MCP payloads of 5MB+.
Next.js frontend
A React/Next.js interface talks to the backend over Server-Sent Events for live streaming — no page reloads, no polling.
Choice of multi-agent framework
The same solution can run on LangGraph, Google ADK, or Microsoft Agent Framework — switchable per deployment, not hard-wired to one vendor's orchestration engine.
Agent2Agent (A2A) protocol support
Every agent pipeline also exposes a standards-based Agent-to-Agent (A2A) endpoint with a real, discoverable agent card — so another agent system can call it directly, not just this UI, including mid-conversation clarification round-trips.