Brandon Dent, MD
Speak into the mic at https://prism42-app.thegoatnote.com/prism42/livekit and you reach a real-time 911 voice dispatcher that responds in under 1.5 seconds and walks a full emergency protocol: address verification, compl
Core trick
A deterministic state machine delivers life-safety instructions verbatim while a parallel LLM critic scores each reply for intent agreement.
Description
Speak into the mic at https://prism42-app.thegoatnote.com/prism42/livekit
and you reach a real-time 911 voice dispatcher that responds in under 1.5 seconds and walks a full emergency protocol: address verification, complaint-specific triage, and T-CPR pre-arrival instructions.
This project began by stress-testing Claude Opus 4.7 on HealthBench Hard and MedAgentBench, then building a custom evaluation harness to exceed native model behavior—turning a general-purpose LLM into a safety-constrained, real-time clinical decision system.
Prism42 developed out of a need for 911 call centers to have AI assistance during downtime. A deterministic finite-state machine owns every safety-critical decision and template — CPR instructions, repositioning, breathing verification — delivered verbatim.
A parallel Opus 4.7 critic scores intent agreement across a 100-fixture evaluation. Naive LLM dispatchers drift and improvise; Prism42 supervises with an LLM and dispatches with an FSM.
A cloud demo (https://prism42-console.vercel.app/prism42-v3
) uses ElevenLabs + Claude for smooth playback; the full system runs self-hosted on a Brev B300 (https://prism42-app.thegoatnote.com/prism42/livekit ).
Every life-safety path is physician-reviewed and traceable to NHTSA EMD, AHA T-CPR, and NHS Pathways. 83/83 tests pass. MIT licensed.
When most AI systems go silent, Prism42 keeps responding.
README excerpt
<p align="center"> <img src="assets/prism42-medical-rag.png" alt="prism42 medical RAG · all-GPU runtime — Refract complexity. Ship the answer. User medical inquiry → NVIDIA knowledge-graph RAG → B300 all-GPU native (nx-cugraph in-VRAM graph + Nemotron Nano 30B + Cosmos vision) → five adversarial roles → safe final response, with a Karpathy autoresearch nightly feedback loop on retrieval/ranking/subgraph logic." width="100%"> </p> # Prism42 > **Maintenance status (2026-09):** passive. This repository is kept available as a reference implementation; CI runs on pushes and pull requests only, Dependabot security alerts remain enabled, and no scheduled jobs or hosted services consume ongoing resources. No active development is planned. **A full-stack trust-and-performance pipeline for high-stakes voice AI. Find correctness failures. Optimize the compute path. Prove clinical- reasoning lift. Deploy the agent stack. Package into a 911 call-center demo anyone can interact with.** Prism42 proves three things before the demo runs: the agents are **correct**, they are **fast**, and they are **clinically safer than baseline**. The pipeline is four stages composing into one deployable system — see [`docs/pipeline-narrative.md`](docs/pipeline-narrative.md) for the full thesis. No speculative findings. No benchmark numbers we didn't measure ourselves. Every claim on the landing page traces to a session ID; every session ID reproduces with a shell command; every agent the public tal