SwiftCODE
SwiftCODE is a hands-free, edge-native agent for code blue and rapid response events.
Core trick
Local speech transcription feeds deterministic rules that timestamp callouts as events, and low-confidence ones need one-tap confirmation.
Description
SwiftCODE is a hands-free, edge-native agent for code blue and rapid response events. During a resuscitation, clinicians already call out medications, shocks, rhythm checks, airway placement and other interventions, but a recorder nurse must still capture those events manually under intense pressure. Paper documentation can be incomplete and timing can be inaccurate.
SwiftCode listens to the room, transcribes speech in real time, extracts and timestamps clinical events, and runs advisory ACLS timers for CPR cycles and epinephrine intervals. The critical loop runs locally using SwiftInference, so transcription, event logging and timers continue even when hospital connectivity fails. Deterministic rules handle clear callouts instantly, while a local language model resolves more complex speech. Low-confidence events require one-tap human confirmation.
At ROSC or termination, Claude converts the verified event timeline into a professional code record and plus/delta debrief, with structured JSON and Markdown export for downstream documentation platforms.
SwiftCODE does not diagnose rhythms or direct treatment. It is a human-in-the-loop documentation and protocol-timing support tool designed to reduce recorder burden, improve timestamp accuracy and create a usable record of one of healthcare’s most chaotic moments
README excerpt
# SwiftCode A hands-free, edge-native agent for code blue and rapid response events. It listens to a resuscitation room, transcribes speech in real time, extracts and timestamps every intervention, runs ACLS protocol timers, prompts the team when actions are due, and generates a timestamped code record and debrief at ROSC or termination. SwiftCode is a documentation and protocol-timing support tool with humans in the loop. It never interprets ECG rhythms and never directs therapy autonomously. It supports data locality requirements relevant to healthcare; it does not claim regulatory compliance, and it makes no claim of survival benefit or clinical outcome. Built for the Abridge x Anthropic x Lightspeed hackathon (SwiftInference, July 18 2026). See [CLAUDE.md](CLAUDE.md) for the full build specification. ## Architecture The real-time loop (speech-to-text, event extraction, timers, prompts) runs entirely on the local/edge node with zero internet dependency. A deterministic rules layer (`src/pipeline/rules.js`) catches clean ACLS callouts in under a millisecond; a local LLM via the SwiftInference API runs behind it only to disambiguate messy speech, and the real-time loop survives total LLM or network failure. Claude is used only for the post-event summary (narrative code record + plus/delta debrief), never in the real-time loop, and its calls queue automatically when offline. ## Setup Requires [Bun](https://bun.sh). ``` cp .env.example .env # fill in SWIFT_API_KEY and