Skip to Main Content

LightDay

Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

LightDay — Demo video

# Lightday — One Care Engine. Three People. One Patient Timeline. Lightday connects three key roles through an intelligent clinical engine: * **Doctor (Dr. Rao):** Dictates visits, medications, and discharge summaries. The chart writes itself; the physician decides and signs. * **Nurse (Asha/Meena):** Photographically verifies wristbands, vitals, and consent. Camera tracking counts instruments/swabs, blocking closure until saved. * **Patient (Vani):** Checks in remotely, matches medicine strips to prescriptions, and accesses medical records instantly. --- ### Three Roles. Three Meta Models. | Role | Model | Core Responsibility | | --- | --- | --- | | **Watcher** | SAM 3.1 | Counts instruments & swabs (tray to bucket) | | **Listener** | Muse Voice | Transcribes dictation, orders & medications | | **Reader** | Muse Spark 1.3 | Audits wristbands, vitals, strips & drafts | > **Core Rule:** AI never decides; a human confirms. > Every prediction provides an audit view: raw input, output, usage, and clinician actions. --- ### Impact (30%) Lightday eliminates retained surgical swabs, cuts medication errors, and replaces paper tracking across clinics. **Proven Workflows:** * **15/15/0 Count:** Camera-based swab/instrument count completed and verified. * **Medication Match:** Home blister strips matched to electronic prescriptions. * **Safe Discharge:** Unlocked only after all clinical checklist evidence aligns. Two production modules (**Doctor Dictation** & **Ward Care**) share one patient timeline, one Meta Model API key, and an inspectable audit log. --- ### Meta AI Tooling (25%) Meta models serve as active senses, not generic chatbots: * **SAM 3.1 (Vision):** Audits pre/post-op surgical counts and identifies missing items via frame evidence. * **Muse Spark 1.3 (Vision & Reasoning):** Verifies wristbands, vitals, and consent (5.9s–9.7s latencies); reads packaging and drafts handovers. * **Muse Voice (Speech):** Captures operative orders and bedside notes into structured entries requiring confirmation. *Telemetry logs execution step, latency, and failure states. Rules-only outputs are explicitly marked.* --- ### Technical Depth (20%) & Demo (25%) Monorepo with unified `.env` (`META_API_KEY`): * **Doctor Dictation:** `doctor-dictation/` (port 8000) * **Count Guard:** `count-guard/` (frontend 5173, API 8100) **Highlights:** * Flagged count mismatches block surgical closure; invalid photos are rejected with reasons. * Non-destructive corrections preserve provenance. Tested via `pytest`, `node`, and `make demo-check`. * 3-minute video built from 18 live-API captures, voiceover, and a MusicGen score. Local switcher dock navigates Doctor, Nurse, and Patient roles seamlessly. --- *Safer, connected clinical workflows via multimodal AI — keeping every clinical decision in human hands.*

Team