OpenAI Build Week

OpenAI Build Week · Work & Productivity · 2nd place

Pulse

Built by Mohamed Mostafa

A cardiac-arrest research prototype with a deterministic engine for rhythm, shocks, and drug timings. Our runs used simulated transcription; live room-audio recognition was not exercised.

I also participated in this hackathon with Trajectory. This coverage was prepared independently after results were announced.

The idea worth stealing

The confirmation step was already in the room.

Resuscitation teams close the loop out loud: “shock delivered,” “epi is in.” Pulse treats that spoken completion as the acceptance, keeps a 30-second undo, and leaves orders pending until the loop closes, so nobody touches a screen. The AI’s job shrinks to hearing; the protocol does the deciding. That is why the engine can be fully deterministic and still feel like a teammate.

What we checked

Date checked: .

Submission code and tests with fake speech recognition; no live room-audio test.

Labels apply to each finding and the evidence described, not to the project as a whole.

Not independently verified means the available evidence was insufficient to confirm a claim. It does not mean the claim is false.

  1. No language model in the clinical loop, or anywhere else

    Verified

    The clinical engine imports nothing from any AI module, and eight boundary tests keep the copilot from touching state. There is in fact no LLM in the shipped runtime at all; the only model calls are speech-to-text. GPT-5.5 and 5.6 were the Codex build tools. Source

  2. It runs end to end with no API key

    Verified

    We ran the scenario sweep in an empty environment. All 29 scenarios pass, including code-switched Egyptian Arabic, echo duplicates, and a shock spoken during asystole. So do 443 backend and 39 frontend tests. Source

  3. Silence fails closed

    Verified

    The browser drops any audio segment quieter than −40 dBFS before upload, and the server treats an empty transcript as an error, not an event. Echoes are deduplicated per event type: 15 seconds for shocks, 10 for CPR and rhythm. Source

Under the hood

Evidence, not facts
Audio becomes a transcript, then a normalized observation, then evidence with a confidence, then a candidate event. Only a stated completion auto-accepts; an order waits as a pending card. Every acceptance and undo is written to the audit log.
Speech-to-text is a swappable vendor
One adapter reads a vendor, model, and key from the environment: OpenAI gpt-4o-transcribe, Groq Whisper, ElevenLabs Scribe, or the deterministic fake. Each is primed with a resuscitation vocabulary so the recognizer expects the right words.

Nuance

Everything we ran used the deterministic fake ASR; live transcription of noisy room audio was read, not exercised, and the demo film is the only live recording we saw. The submission says “the LLM helps read messy speech,” but in the shipped code that reading is a speech-to-text model plus hand-written phrase rules. And the README is blunt: research prototype, not a medical device.

Sources

Last updated: . Editorial updates do not imply a new technical check.

Builder credits reviewed: , using the official announcement, submission, and any linked credit sources.

Builder? Add context, request a correction, or ask for a re-check →