Inspiration: We've all seen the chaos of an emergency handoff — information passed verbally under pressure, details dropped, the receiving team left guessing. One of our teammates had a family member go through a traumatic injury, and watching the EMS crew work, we noticed how much coordination happened in people's heads with no record. When we found research showing that 43% of critical EMS handoffs were missing a complete set of vital signs, we knew the problem was real and widespread. We wanted to build something that actually lives in the workflow — not a clipboard, not an app someone has to open.
What it does: Relay is a voice-controlled coordination system for EMS crews. Each crew member wears a VexoBand — a wristband with a microphone, LED, and tap sensor. When the crew lead assigns a task by voice, the right medic's band vibrates privately. That medic double-taps to acknowledge. When they complete the task, they speak the result. At every step, Relay tracks the full lifecycle: requested, acknowledged, completed, verified. At hospital handoff, Relay reads out a structured summary that separates confirmed findings from tasks that are still open — so the receiving team knows exactly what happened and what hasn't been resolved yet.
How we built it: We built Relay on three connected layers. The hardware layer uses VexoBands connected to laptops over Bluetooth via the Vexo Open Kit. A Python bridge script runs for each crew member — it listens for single-taps to toggle the microphone, streams audio to Deepgram for real-time transcription, then sends the transcript to OpenAI to classify the intent: is this a task assignment, a completion report, a correction, or ambient noise? The result is posted to a central Express/TypeScript server that maintains the task ledger and pushes live LED commands back to the right band. Confirmations and handoff summaries are spoken aloud using MiniMax text-to-speech. A live dashboard streams every state change in real time so anyone nearby can see the board.
Challenges we ran into: The hardest problem was the microphone feedback loop. When Relay speaks a confirmation aloud, the band's microphone picks it up, Deepgram transcribes it, and OpenAI tries to dispatch it as a new command — creating phantom tasks. We solved this by broadcasting speaking state over the server's SSE stream so the band bridge suppresses transcript dispatch while audio is playing. The second challenge was the push-to-talk flow: a double-tap on a real band fires two single-tap events first, which would accidentally toggle the mic before the acknowledgment could be posted. We added a pending-flag pattern in the main loop so mic toggles are deferred and can be cancelled cleanly by a double-tap.
Accomplishments that we're proud of: We're proud that Relay actually works end-to-end in a live demo — voice in, band vibration out, spoken confirmation back, task board updating in real time. The handoff output — with confirmed events and unresolved tasks cleanly separated — is the thing we're most proud of. It's not a transcript. It's a structured record of what happened and what didn't, generated automatically from the crew's normal conversation.
What we learned: We learned that the hardest part of building voice systems isn't the speech recognition — it's everything that happens around it. Suppressing feedback, handling partial transcripts, deciding when to dispatch and when to wait. We also learned that the framing of the problem matters as much as the solution: Relay isn't an AI scribe, it's a coordination tool that generates documentation as a side effect. Getting that distinction clear changed how we designed every feature.
What's next for Relay: We want to test Relay with a real EMS crew. The immediate next step is a structured pilot — two medics, two bands, a simulated call — to see how the haptic vocabulary holds up under pressure and whether the intent classification is reliable enough for clinical use. Longer term, we want to close the loop on corrections and corrections history, build out the evidence-linked timeline so every handoff statement traces back to its source, and explore integration with existing EHR handoff workflows.
