About the project
Prior authorization is a protocol problem wearing a paperwork costume.
Pavo Cloud was built for the AI2O finals as a working system rather than a pitch — every claim on this site can be traced to code in the repository or to a measured run.
The thesis
The data already exists. The two sides just cannot talk.
When a physician orders an MRI, the clinical record already contains every fact the payer will ask for. The delay does not come from missing information. It comes from the absence of a channel: the provider's system and the payer's system have no way to exchange a question and an answer, so a human transcribes one into a fax and another transcribes it back out.
Three to fourteen days and $35 billion a year go into that transcription. For the cases that are clear-cut — and most are — no judgement is being exercised at any point in the chain.
So the problem is not that the decision is hard. It is that there is no protocol. Build the channel, make both ends prove who they are, and the clear cases settle themselves.
The approach
Why this is not a language model deciding coverage.
Decisions are rules
Coverage is a contract. A model that is right 95% of the time is not acceptable when the other 5% is someone's surgery, and it cannot tell you which rule it applied. Pavo's engine cannot return an outcome without naming one.
Identity is cryptographic
Two organizations that have never met need more than a shared API key. Every message is signed, and an unverified one is rejected before its payload is read.
Privacy is structural
Rather than sending the chart and promising to be careful with it, the provider proves the criteria hold and sends nothing else.
Models do the work models are good at — reading an insurance card, mapping a plain-language procedure to a CPT code, drafting an appeal letter from the clinical record. None of them decides whether a patient is covered.
Where it stands
Built, next, later.
Stated at the level of detail a reviewer would want, including the parts that are not done.
Built
The authorization loop
Signed agent-to-agent messaging, a deterministic rule engine, zero-knowledge proofs over patient criteria, autonomous appeals, and an append-only audit trail.
Next
Real pricing, real registries
Replace the deterministic price generator with facility rates gathered by the voice agent, and turn NPI verification from a flag into a lookup against the live CMS registry.
Later
A circuit that hides the set
A Merkle circuit, so a proof shows the diagnosis is one of a payer's covered conditions without revealing which one, plus a real multi-party trusted setup.
Later
Federated learning
Described in the specification and not built. Payers would improve a shared model of what escalates without raw records leaving any payer's node.
6 subsystems run today; 5 known limitations are listed in full on the home page and in the repository README.
The fastest way to judge it is to run it.
The demo executes against the same agents, rules, signatures and circuit as everything else here. Nothing on that page is a recording.