RxRelay
Maya is a consent-first voice coordinator for prescription access — a case closes only when the evidence exists.
The problem
Prior authorization stalls. Inventory shifts. A form never arrives. The patient becomes the switchboard — calling pharmacy, clinic, insurer, repeating context, and still getting no answer they can trust. Voice agents look like the fix. The failure mode is subtler: an agent that says “I’ve taken care of it” and an agent that actually coordinated something look identical at the transcript layer. In medication access, that gap is the whole risk.
PAVO routing
Grounded in PAVO: Pipeline-Aware Voice Orchestration with Demand-Conditioned Inference Routing (VeiluKanthaPerumal & Imthathullah, TMLR 2026). The paper’s coupling cliff is empirical: a stronger LLM cannot repair a misheard authorization number once upstream ASR has already lost the fact. RxRelay implements that constraint in production voice turns — without pretending to swap between fictional multi-vendor ASR backends.
On the live TeXML path, PAVO selects a joint capture + reasoning tier.
Verified turns widen the a1mobile <Gather> input from speech-only to
speech dtmf, add digit capture for critical references, and route inference to a stronger
reasoning model with structured confirmation — together, never one without the other.
There is no runtime hop between competing cloud ASR vendors. Tier labels (edge_fast,
hybrid_balanced, cloud_premium) map PAVO paper routes onto
TeXML speech settings + model class + confirmation policy inside one orchestrator.
The upgrade is real; the vendor carousel is not.
Low-demand intake — speech gather, compact reasoning, no protected action.
Coordination turns — hybrid capture, tool-aware reasoning, permitted status actions only.
Coupling-cliff risk — speech + DTMF, strong reasoner, confirm-before-act on names, dates, auth numbers.
Clinical advice, emergencies, Rx changes, controlled inventory — human handoff, zero autonomous outreach.
Paper: openreview.net/pdf?id=zrneoIxlFx · Benchmark: vnmoorthy/pavo-bench · Implementation: src/pavo.mjs
Proof gate
Resolution is a pure function over recorded state. Four checks must pass before status becomes
resolved — consent, permitted action, counterpart outcome, patient notification.
Partial progress stays open. A failed provider call records no action evidence.
A clinic submission does not imply a ready pharmacy pickup.
Product surfaces
RxRelay is not a transcript viewer. It is a consent-gated coordination ledger — natural inbound voice, operator tooling, counterpart attestation, and live visibility — all wired to the same proof gate. Live outbound telephony stays intentionally fail-closed until OTP allowlist and provider confirmation.
Operator dashboard at :3000. Pick a case, walk the sandbox coordination path, watch the
close gate flip only when evidence exists. PAVO lab exposes demand scores and route decisions per turn.
Magic-link attestations for pharmacy, clinic, or insurer staff. Third parties confirm blockers, PA submission, or readiness — replacing patient-spoken stand-ins with recorded counterpart facts.
Hash-chained JSON receipts on case close: consent text, provider action IDs, route decisions, event chain. Pure Node crypto — verifiable without trusting model narration.
Safe-stop and escalation paths set humanReview and block autonomous completion.
Urgent medical cues, clinical advice requests, and ambiguous outcomes route to a human owner — not deeper automation.
Server-sent events stream case mutations to the proof board in real time. Voice turns, webhook events, and coordination actions appear without polling — the operator sees the same state the gate evaluates.
Architecture
Two processes share one case file. The Cloudflare quick tunnel reaches only
voice-server.mjs — token-gated TeXML on :3001, no dashboard, no MCP, no
outbound credentials in the voice request path. Dashboard, webhook seam, and MCP live on
server.mjs at :3000. Shared state persists through
data/cases.json via persist.mjs.
TELEPHONY_PROVIDER=demo — full E2E proof path with zero runtime dependencies. No real dial or SMS.
Real TeXML speech loop via claimed a1mobile number. Inbound voice ≠ outbound messaging until OTP allowlist + provider action id.
Live adapter refuses completion without provider-issued action id. Failed calls never manufacture action evidence.
Run it
Clone, copy .env.example, run tests — no npm install.
The sandbox walks a case from consent through pharmacy blocker, clinic submission, and verified close.
Live inbound is a configuration decision: tunnel, point the number, speak naturally — outbound stays gated until you explicitly enable it.
git clone https://github.com/vnmoorthy/rxrelay.git
cd rxrelay
cp .env.example .env
npm test # 27 tests · node:test · zero deps
npm run deck # PPTX → deck/output/… ; HTML at deck/pitch.html
npm run dev # proof board → localhost:3000
npm run live:inbound # Maya TeXML :3001 → tunnel → point numberOn a live call, say something like “I need help with my prescription — the pharmacy says prior auth” — Maya records scoped consent, walks the permitted path, and keeps the case open until counterpart evidence lands.
Why a1mobile should invest
a1mobile’s stack already carries the calls. RxRelay shows what happens when TeXML speech, webhooks, and OTP-gated outreach feed a case engine that refuses to hallucinate completion. The wedge is prescription access — high-frequency, high-anxiety, coordination-heavy — where proof beats polish.
TeXML <Gather> speech and DTMF, number claim/point scripts, SMS OTP verification,
webhook events — not a parallel telephony abstraction. PAVO upgrades map directly onto capture mode changes
in the live voice loop.
Every competitor demo ends at “the agent said done.” RxRelay exports hash-chained receipts, counterpart attestations, and a deterministic gate that partners and compliance teams can audit. That is a platform story, not a prompt-engineering story.
PAVO is peer-reviewed (TMLR 2026) with an open benchmark. RxRelay is the reference implementation for clinical-ops voice — demand-conditioned, coupling-aware, honest about what upgrades and what does not.
Zero runtime dependencies. CI green. Sandbox E2E proven. Live inbound path exercised. Live outbound fails closed until OTP allowlist and provider action confirmation — the safe default for a Catalyst investment conversation.