Praxess
Prior authorization delays care because the evidence needed for approval is scattered across the clinical conversation, note, FHIR record, payer policy, patient, and outside providers.
Core trick
Verifies each quoted transcript span against source text before it enters the case state, tagging each fact with its provenance tier.
Description
Prior authorization delays care because the evidence needed for approval is scattered across the clinical conversation, note, FHIR record, payer policy, patient, and outside providers. Key facts get lost between what the patient says and what enters the chart. Teams discover those gaps after submission, when the payer requests more information or denies the case.
Praxess is a world-model prototype for prior authorization. It maintains a living case state from the moment the patient enters the room, with source provenance attached to each claim.
Claude structures evidence from the transcript and separates doctor and patient turns. Praxess then verifies quoted spans against their source before adding them to the case state. It tracks the difference between clinician-documented, conversation-enriched, patient-reported, verified, and unknown information.
The decision engine rolls candidate actions forward and scores them using expected approval lift, information gain, time cost, and staff burden. A person approves consequential actions. Each new observation updates the state and changes the recommendation.
Praxess records reviewed state-action transitions, creating the foundation for learned dynamics as authorization outcomes accumulate.
The demo follows one authorization from encounter capture through evidence recovery, patient outreach, external record retrieval, packet generation, submission, denial, and appeal.
If you want to check it out yourself, here's the link: https://praxess-production.up.railway.app/
README excerpt
# Praxess **Prior auth from the room** — a closed-loop prior authorization agent that mines the clinical conversation for medical-necessity evidence, verifies every claim deterministically, drives the case to approval with a clinician approving every consequential step, and keeps running after a denial. > Built during the **Abridge × Anthropic × Lightspeed HealthTech Hackathon** (July 18, 2026). > Synthetic organizer-provided data only. --- ## The problem in one sentence The evidence for most PA denials was **spoken in the visit** — the note just lost it. A physician says: *"You mentioned ibuprofen for a while — did that help?"* The patient answers: *"Yeah, worked great, but the bottle ran out months ago."* The note records: *"Ibuprofen prescribed."* The payer sees: no documented trial duration → denial. Praxess listens to that conversation, finds what the note missed, prices the next best action, drafts the fix behind a clinician approval gate, and — when the payer still says no — plans the appeal from the same evidence. --- ## Live demo **[praxess-production.up.railway.app](https://praxess-production.up.railway.app)** --- ## Demo one-liner *The evidence for this denial was spoken in the visit. The note lost it. We found it, priced the next best action, drafted the fix, and the doctor approved it in one click — and when the payer says no, the same loop plans the appeal.* --- ## Quick start ```bash ./scripts/dev.sh # venv + npm install + API on :8000