Google Gemini Live Agent Hackathon

Voice-Directed Intelligence
for the Operating Room

ORION gives surgeons hands-free access to everything — patient data, CT imaging, 3D anatomy, and AI-driven surgical checklists — using only their voice, while their hands never leave the controls.

"ORION, show me the hemoglobin"
da Vinci Surgical System
da Vinci Xi System

img source: ohcare.com

The Problem

Locked hands.
Critical information
out of reach.

During robotic surgery, the surgeon's hands are locked on instrument controls inside a sterile field. They cannot type, click, or interact with any computer system.

Critical patient data, imaging, and anatomical context require them to break scrub or rely on circulating staff — both slow, disruptive, and potentially dangerous at the wrong moment.

Sterile Field Constraint
Hands never leave the controls during critical phases
Information Latency
Relying on staff to retrieve data adds cognitive load and delay
Context Switching Cost
Breaking scrub for data interrupts surgical flow and focus
Challenges ORION addresses
CVS Documentation
Only 23.1% of laparoscopic cholecystectomies have Critical View of Safety documented — the primary safeguard against bile duct injury.
Orion's Safety Protocol Agent logs CVS confirmation by voice and immediately captures a timestamped photo as a medicolegal record.
Terho et al. 2021 · PMID 33975802
WHO Surgical Checklist
Implementing the WHO Surgical Safety Checklist reduced in-hospital mortality by 47% and complications by 36% globally.
Orion's Timeout Agent surfaces the phase-specific WHO checklist on screen and marks it complete once the surgeon confirms each item by voice.
Haynes et al. 2009 · NEJM · PMID 19144931
Blood Loss Estimation
Surgeons underestimate intraoperative blood loss by 52–85% on average; 95% of clinicians are wrong by more than 25%.
Orion's Blood Loss Agent tracks a running total against the patient's estimated blood volume and escalates alerts at 15%, 25%, and 40% loss thresholds.
PMC7943515
Operative Note Delays
Surgeons take a mean of 15.6 days to dictate operative notes — vs. just 28 minutes with structured voice templates.
Orion's DOC Agent auto-timestamps every milestone throughout the case, producing a complete structured operative record the moment the case closes.
Laflamme et al. · PMC1560865
Drug Administration Errors
1 in 20 anesthesia drug administrations involves an error — 80% are judged preventable with better real-time cross-checks.
Orion's Drug Safety Agent cross-checks the patient's allergies, active medications, and live lab values — returning a safe or caution verdict on screen in under a second.
Nanji et al. · PMC4681677
How It Works

Speak. Process. Render.

PIPELINE
Voice to Surgical Display
Hover any step below to trace its path through the live pipeline.
24 TOOLS SURGICAL DISPLAY SURGEON VOICE GEMINI LIVE API · ORION ORCHESTRATOR Pre-OpBriefing SurgicalTimeout OperativeReport SafetyProtocol BloodLoss Est. DrugInteraction AnatomyNavigator HandoffSBAR ScreenAdvisor
01
Voice_Cmd
Voice Command
Surgeon speaks. 16kHz PCM audio streams via WebSocket. Wake-word filtering ignores ambient OR noise and staff conversation.
16kHz PCM Wake-word
02
Gemini
Gemini Live API
Audio and live surgical video (1 fps) processed natively. Full multimodal context — the AI understands the OR environment in real time.
Native audio Live video
03
Routing
Multi-Agent Routing
ADK Orchestrator routes the intent to the right specialist — patient data, CT imaging, 3D anatomy, surgical phase, or intraoperative documentation.
ADK AutoFlow 5 specialists
04
Render
Visual Overlay
Results appear as tiles on the surgical display. ORION speaks a brief confirmation. Hands never leave the controls.
Tile layout Hands-free

One orchestrator.
Nine specialists.

ORION routes every voice command to the right specialist agent — protocol, clinical decision, documentation, or screen analysis — no configuration, no delays, no hands.

ROOT AGENT
ORION Orchestrator
Always-on. Listens for the wake-word, filters ambient OR noise, and routes each command to the right specialist via ADK AutoFlow. Handles eight specialists — pre-op briefing, WHO timeout, operative report, complication response, blood loss tracking, drug safety, anatomy guidance, and patient handoff. Never announces routing.
Wake-word filtering AutoFlow routing Multi-turn 8 specialist agents 24 tools
001
Briefing_Agent
Pre-Op Case Briefing
Delivers a concise pre-operative briefing — patient identity, key labs, allergies, and phase-one checklist highlights. Ready in seconds.
Labs & vitals Allergies Phase briefing
002
Timeout_Agent
WHO Safety Timeout
Runs the WHO Surgical Safety Checklist at incision. Confirms patient identity, procedure, allergies, and critical labs — hands-free and timestamped.
WHO checklist Patient ID confirm Allergy check
003
Report_Agent
Operative Report Generator
Compiles the intraoperative event log into a structured operative report. SBAR-style findings exported to the surgical display instantly.
Auto-documentation Event log SBAR format
004
Complication_Advisor
Complication Protocol Agent
Responds to intraoperative complications — bleeding, nerve injury, air leak, conversion. Surfaces the phase-aware management protocol instantly.
Bleeding response Conversion protocol Phase-aware
005
EBL_Tracker
Blood Loss Tracker
Tracks cumulative estimated blood loss from voice updates. Maintains a running total and alerts when thresholds approach transfusion criteria.
EBL tracking Running total Transfusion alert
006
Drug_Checker
Intraoperative Drug Safety
Real-time drug safety check against the patient's medication list and allergy profile. Flags contraindications and interactions before administration.
Allergy check Drug interactions Pre-dosing
007
Anatomy_Spotter
Anatomy Identification Agent
Identifies at-risk structures for the current surgical phase. Uses 3D model context to surface danger zones and critical spatial relationships.
Danger zones Structure ID Phase-aware
008
Handoff_Agent
Patient Handoff — SBAR
Generates a structured SBAR handoff for shift change or scrub-out. Pulls the event log, phase status, and patient data automatically.
SBAR format Shift handoff Auto-summary
009
Screen_Advisor
Visual Console Intelligence
Captures a live screenshot of the surgical console and uses Gemini vision to answer questions about what is visible — panels open, vitals shown, CT slice displayed.
Screen capture Vision analysis Live context

How ORION Works Under the Hood

End-to-end flow from surgeon's voice to real-time surgical intelligence — ADK multi-agent orchestration, Gemini Live API, and a WebSocket audio pipeline.

ADK Features Used
LlmAgent with sub_agents
transfer_to_agent routing
before / after tool callbacks
Runner.run_live() BIDI
LiveRequestQueue
aclosing() cleanup
Gemini Live Features
Native audio dialog model
BIDI streaming (2-way)
Real-time audio transcription
PrebuiltVoiceConfig
Function calling in live
Multi-agent audio generation
Architecture Patterns
Per-session async WS proxy
AudioWorklet → PCM pipeline
100ms audio chunking
asyncio.wait FIRST_EXCEPTION
camelCase JSON aliasing
Real-time event streaming

Built on Gemini Live

A production-grade stack using Google's latest multimodal real-time AI, purpose-built ADK agents, and a high-performance WebSocket server.

Gemini Live API
Native audio + vision, real-time
Google ADK
Multi-agent orchestration
FastAPI + WebSocket
Bidirectional streaming server
Three.js
Real-time 3D anatomy renderer
AudioWorklet API
16kHz capture / 24kHz playback
LIDC-IDRI CT Data
133-slice DICOM, GCS-hosted
Vertex AI
Gemini model serving
Python 3.11
Async server, ADK runners
Live Demo

Experience ORION
in the operating room

Connect your microphone, speak a command, and watch ORION route it to the right agent — all in real time.

Click the orb to connect · Say "ORION, show hemoglobin" to begin

This project is designed to demonstrate the capabilities of Gemini Live API & Google ADK. It may contain clinical inaccuracies and has not been reviewed by medical domain experts. Not intended for use in actual clinical settings.