Skip to content
Mohib Ahmad
All work
AI & Data2026 · prototype · partly public

Priora

A provider-side prior-authorization control plane for US healthcare, built on criterion-level evidence rather than denial prediction.

The problem

Prior authorization is slow, document-heavy, and opaque. Reviewers need to know which criteria are met, on what evidence, under which version of which policy — not a probability that a payer will say no.

For: Provider-side operators, clinic teams, and reviewers assembling authorization packets.

What I built

A prototype with a Next.js frontend, an API backend, server-side document extraction, model-assisted analyzis, and optional Postgres persistence. The design work that followed it is larger than the prototype: a criterion DSL, a five-value status algebra with explicit applicability, a canonical state machine, an event model with hash-chained audit, and a determinism boundary that keeps models out of every verdict.

The hard part

Deciding that the original output was the wrong output. The prototype predicted denial risk; the architecture that replaced it forbids a denial probability from sorting a work queue or feeding any gate, because a probability tells a reviewer nothing they can act on or defend.

The tradeoff

Chose evidence and provenance over the demo-friendly prediction score. It is a harder system to build and a less impressive thirty-second demo, and it is the one a compliance reviewer can actually accept.

What went wrong

The resume and the product pack describe two different products under one name. That is a real inconsistency, and it is recorded rather than smoothed over: the resume describes the prototype, and this page describes what the design became.

What this does not do

  • The public repository is the earlier prediction-era prototype. It does not implement the evidence-first architecture described here.
  • The specification packs are private, so most of this work cannot be independently verified from the public repo alone.
  • No production deployment, no real payer integration, and no clinical validation. Nothing here should be read as a claim of accuracy on real cases.
  • Fifteen open questions in the product pack remain open, deliberately — they are recorded as spikes rather than answered by invention.