Powered by Google Gemini

Helios
Sees Everything.

Multimodal AI assistant for datacenter engineers. Real-time visual reasoning through Meta Ray-Ban smart glasses, grounded in your infrastructure APIs. Zero hallucination.

See It In Action View Source
1fps
Video Analysis
116+
Devices Tracked
<100ms
Voice Latency
0
Hallucinations
Core Capabilities

What engineers
actually need.

Helios combines real-time computer vision, voice interaction, and datacenter API integration into a hands-free assistant that rides on your face.

Visual Reasoning

Sees through Meta Ray-Ban glasses at 1fps. Identifies servers, reads LEDs, recognizes cable types, and inspects rack health in real-time via Gemini 2.5 multimodal.

Voice-First Interface

Techs have their hands full. Helios speaks and listens through the glasses -- no phone, no screen, no extra hardware. Natural conversation with automatic speech detection.

NetBox Integration

Pulls live device inventory, rack positions, cable maps, and IP assignments from your NetBox instance. Every answer is grounded in your actual infrastructure data.

Redfish Health Monitoring

Real-time CPU temperatures, fan speeds, power consumption, and system health from Redfish BMC APIs. Correlates visual cues with telemetry data for accurate diagnosis.

State Machine Tracking

Continuous JSON state engine tracks task progression, urgency levels, and time-to-action predictions using linear regression. No context is ever lost between frames.

WebRTC Live POV

One-way video stream from glasses to a browser dashboard via WebRTC. Remote teams see exactly what the technician sees with room-code based signaling.

System Design

End-to-end
architecture.

From photons hitting the glasses camera to voice response in the technician's ear -- every hop in under 100ms.

Smart Glasses
Meta Ray-Ban Camera + Mic
Helios iOS
Swift + Meta DAT SDK
Gemini 2.5
Multimodal WebSocket
DC APIs
NetBox + Redfish
Voice + Alerts
Audio Response

iOS App

  • SwiftUI with AVAudioEngine pipeline
  • Meta Wearables DAT SDK for glasses streaming
  • WebSocket client for Gemini real-time API
  • 16kHz mono input, 24kHz playback
  • JPEG frame encoding at 50% quality, 1fps

AI Layer

  • Gemini 2.5 Flash native audio preview
  • Multimodal: video + audio + text context
  • Structured JSON state output per response
  • Tool calling for agent gateway delegation
  • Automatic speech detection with interrupts

Infrastructure

  • NetBox API for device inventory & topology
  • Redfish BMC for real-time server health
  • 30-second refresh cycle for health data
  • Mock data generators for demo scenarios
  • 3 sites, 116 devices, full rack mapping

Streaming

  • WebRTC one-way video (glasses → browser)
  • Node.js signaling server with room codes
  • TURN server for NAT traversal
  • 60s grace period for backgrounding
  • Custom video capturer for WebRTC pipeline

Critical alerts, delivered
before it's too late.

Helios continuously correlates what the camera sees with live telemetry from Redfish and NetBox. When a server overheats, a fan fails, or a cable is misrouted -- the technician knows immediately through voice and visual alerts.

Temperature Spikes

Urgency escalates linearly; immediate voice alert at 0.9+

Rack-Level Mapping

Exact U-position, device name, and health status per slot

Visual + API Convergence

Camera observations cross-referenced with Redfish telemetry

9:41

Helios

Ashburn DC-01 · Data Center Mode
Gemini
Agent
CRITICAL: Server Overheating 2s ago
dell-r750-ash-03 in Rack A-07, U12 reporting CPU temp 94°C. Fan 3 failure detected via Redfish. Ambient temp elevated. Immediate inspection required.
Rack A-07, U12 Urgency: 0.92 Act now
Rack A-07 · Ashburn DC-01 4/6 Healthy
U14
cisco-n9k-ash-01 42°C
U13
hpe-dl380-ash-01 58°C
U12
dell-r750-ash-03 94°C
U11
dell-r750-ash-02 71°C
U10
dell-r750-ash-01 55°C
U09
What's wrong with this server?
That's dell-r750-ash-03 in U12. CPU temp is 94 degrees -- fan 3 has failed. I recommend swapping the fan module immediately.
Intelligence Layer

Every response carries
structured state.

Gemini outputs a JSON state object with every response. The app's TaskStateManager tracks urgency, predicts time-to-action, and detects acceleration -- all grounded in what the camera actually sees.

datacenter_state.json
{
  "task": "inspect rack A-07 cooling",
  "work_type": "datacenter",
  "stage": "diagnosing_thermal_issue",
  "urgency": 0.92,
  "cues": [
    "amber LED on U12 server",
    "fan 3 audibly slower",
    "heat shimmer above chassis"
  ],
  "action": "replace fan module immediately",
  "seconds_est": 30,
  "confidence": 0.97
}

Urgency-Driven Response

The urgency field (0.0 - 1.0) drives the AI's communication style and the app's alert behavior. Linear regression predicts how fast urgency is climbing.

0.0 - 0.2 Calm

Brief positive confirmation. "All looks good here."

0.5 - 0.7 Elevated

Direct callout with specifics. "U12 is running warm, 71 degrees."

0.9+ Critical

Immediate alert. "Stop -- that server needs to be shut down now."

Operational Domains

Two worlds,
one assistant.

Helios ships with two domain-specific AI configurations. Each domain has its own system prompt, state schema, and urgency model.

Data Center

Infrastructure monitoring and technician guidance powered by NetBox inventory and Redfish health data.

  • Device identification via camera + API lookup
  • LED and cable recognition in real-time
  • Rack inspection with U-position mapping
  • Temperature and power anomaly detection
  • Troubleshooting guided by visual + telemetry convergence
  • 3 sites, 116 devices, full cable topology

Cooking

Real-time sous-chef guidance with visual state tracking, countdown timers, and anti-burn safeguards.

  • Stovetop and ingredient identification
  • Step-by-step guided cooking phases
  • Visual doneness tracking (opacity, browning, smoke)
  • Countdown timer with urgency prediction
  • Hard caps: 20s over-easy flip, 2min well-done flip
  • Immediate action on smoke detection
The Flow

How Helios works.

From putting on the glasses to getting an AI-powered answer -- every step in the pipeline.

Glasses Connect

Meta Ray-Ban smart glasses pair via the Meta DAT SDK. Video streams at up to 720x1280 in HEVC/H.264.

Frames Captured

The iOS app decodes video frames and sends 1 JPEG per second (50% quality) to the Gemini WebSocket alongside 16kHz audio.

Context Injected

Each frame is annotated with previous state JSON, urgency predictions, datacenter inventory, and health telemetry from NetBox and Redfish.

Gemini Reasons

Gemini 2.5 Flash processes the multimodal input and returns audio speech, text transcription, structured JSON state, and optional tool calls.

Voice Response

24kHz PCM audio plays through the glasses speakers. The technician hears the AI's assessment hands-free while continuing work.

State Tracked

The TaskStateManager updates urgency, predicts time-to-action via linear regression, and feeds this back into the next frame's context.