SkyNav Fleet Manager is a Microsoft Fabric workload designed to help operators monitor and manage a fleet of hot air balloons in real time. It brings together live telemetry, operational intelligence, semantic modeling, and conversational AI inside a single Fabric experience, and it is built with the Microsoft Fabric Extensibility Toolkit.
The scenario is a Loire Valley balloon operator that needs to monitor telemetry, understand operational relationships between balloons, pilots, flights, and sites, and ask natural-language questions about the fleet.
The platform is demonstrated end to end. An IoT telemetry simulator generates realistic balloon data such as position, altitude, speed, flight phase, wind conditions, temperature, and propane level. That telemetry is streamed through Eventstream, can trigger operational alerts in Activator, and is stored in Eventhouse, which becomes the analytical backbone for the live dashboard and the agent.
On top of the raw data, SkyNav adds a business ontology and a connected graph linking balloons, pilots, flights, sites, restricted zones, and weather windows. This semantic layer gives structure and meaning to the telemetry and becomes especially important when users interact with SkyNav Agent, a conversational assistant powered by gpt-5.2 in Microsoft AI Foundry and connected to Fabric data plus MCP-backed weather and wind tools.
SkyNavFullTelemetry.mp4
SkyNav.mp4
Open SkyNav demo video · Open full telemetry demo video
- An IoT telemetry simulator that drives the end-to-end demo scenario
- An Eventstream -> Activator -> Eventhouse pipeline with operational alerting, including low-propane scenarios
- A Fabric workload frontend built with React, Fluent UI, and Leaflet
- A backend API for chat, telemetry, ontology, health checks, and webhook handling
- A conversational sidebar powered by
SkyNavAgent - A
gpt-5.2deployment hosted in Microsoft AI Foundry and consumed through the Azure OpenAI-compatible endpoint - KQL-backed fleet queries plus MCP-backed external weather and wind
- An ontology viewer built from the deployed Fabric ontology definition
/api/copilotkit remains available for compatibility and future evolution, but the current primary user-facing chat path is the custom AgentSidebar.tsx posting directly to /api/chat. Within that backend flow, the core fleet tools remain direct, and skynav-wind is now bridged additively through MCP for external weather and wind.
graph TB
subgraph Fabric[" Microsoft Fabric "]
Portal["Fabric Portal"]
Eventhouse[("KQL Eventhouse<br/>telemetry + weather + alerts")]
Ontology["Fabric IQ Ontology<br/>(getDefinition)"]
Activator["Fabric Activator"]
end
subgraph Azure[" Azure "]
SWA["Static Web App<br/>Workload frontend"]
subgraph Frontend[" Frontend "]
MapTab["Map tab<br/>Leaflet + telemetry"]
OntologyTab["Ontology tab<br/>entity graph"]
AgentSidebar["AgentSidebar<br/>headless chat UX"]
end
subgraph Backend[" App Service backend "]
ChatApi["/api/chat"]
TelemetryApi["/api/telemetry"]
OntologyApi["/api/ontology"]
WebhookApi["/api/webhook/activator"]
SkyNavAgent["SkyNavAgent"]
CopilotRuntime["/api/copilotkit<br/>(compatibility path)"]
end
Foundry["Microsoft AI Foundry<br/>GPT-5.2 deployment<br/>(Azure OpenAI-compatible)"]
end
subgraph MCP[" MCP servers "]
WindMCP["skynav-wind MCP<br/>external weather + wind"]
OtherMCP["data / ontology MCPs<br/>(adjacent capabilities)"]
end
OpenMeteo["Open-Meteo API"]
Portal -->|iframe| SWA
SWA --> MapTab & OntologyTab & AgentSidebar
AgentSidebar -->|POST /api/chat| ChatApi
MapTab -->|poll every 5s when active| TelemetryApi
OntologyTab -->|load graph data| OntologyApi
ChatApi --> SkyNavAgent
CopilotRuntime --> SkyNavAgent
SkyNavAgent -->|chat completions + tool loop| Foundry
SkyNavAgent -->|KQL tools| Eventhouse
SkyNavAgent -->|MCP weather / wind| WindMCP
WindMCP -->|HTTP| OpenMeteo
TelemetryApi -->|KQL queries| Eventhouse
OntologyApi -->|POST getDefinition + LRO| Ontology
Activator -->|webhook| WebhookApi
OtherMCP -.->|available, not primary chat path| SkyNavAgent
style Fabric fill:transparent,stroke:#1565c0,stroke-width:2px,color:#1565c0
style Azure fill:transparent,stroke:#e36209,stroke-width:2px,color:#e36209
style Frontend fill:transparent,stroke:#2ea44f,stroke-dasharray:4,color:#555
style Backend fill:transparent,stroke:#e36209,stroke-dasharray:4,color:#555
style MCP fill:transparent,stroke:#8957e5,stroke-width:2px,color:#8957e5
style Portal fill:#1565c0,stroke:#0d47a1,color:#fff
style Eventhouse fill:#1565c0,stroke:#0d47a1,color:#fff
style Ontology fill:#8957e5,stroke:#6e40c9,color:#fff
style Activator fill:#d73a49,stroke:#b31d28,color:#fff
style SWA fill:#2ea44f,stroke:#238636,color:#fff
style MapTab fill:#2ea44f,stroke:#238636,color:#fff
style OntologyTab fill:#2ea44f,stroke:#238636,color:#fff
style AgentSidebar fill:#2ea44f,stroke:#238636,color:#fff
style ChatApi fill:#e36209,stroke:#c24e00,color:#fff
style TelemetryApi fill:#e36209,stroke:#c24e00,color:#fff
style OntologyApi fill:#e36209,stroke:#c24e00,color:#fff
style WebhookApi fill:#d73a49,stroke:#b31d28,color:#fff
style SkyNavAgent fill:#8957e5,stroke:#6e40c9,color:#fff
style CopilotRuntime fill:#8957e5,stroke:#6e40c9,color:#fff
style Foundry fill:#8957e5,stroke:#6e40c9,color:#fff
style WindMCP fill:#8957e5,stroke:#6e40c9,color:#fff
style OtherMCP fill:#8957e5,stroke:#6e40c9,color:#fff
style OpenMeteo fill:#6e5494,stroke:#553c7b,color:#fff
Blue = Fabric/data, green = frontend, orange = backend, purple = AI/adjacent capabilities, red = automation.
The active chat loop is:
AgentSidebar.tsx -> POST /api/chat -> SkyNavAgent -> gpt-5.2 in Microsoft AI Foundry -> direct backend tools + MCP-backed external weather/wind -> response rendered in the sidebar
The backend tool loop currently focuses on:
get_fleet_statusget_balloon_detailget_active_alertsget_weathergeocode_locationget_external_weatherget_wind_at_altitudeget_wind_forecast
The ontology tab does not depend on undocumented entity endpoints. The backend:
- acquires a backend token with Managed Identity
- calls
POST /ontologies/{id}/getDefinition - polls the long-running operation
- decodes the returned definition parts
- normalizes entities and relationships for the frontend graph
The map is backed by KQL telemetry and trajectory queries. The current frontend behavior is optimized to reduce UI contention:
- telemetry polling runs every 5 seconds only while the map tab is active
- returning to the map triggers an immediate refresh
- full fleet payloads are no longer mirrored into co-agent state every 5 seconds
- the agent sidebar is decoupled from the large fleet context to reduce unnecessary rerenders
| Area | What it does now |
|---|---|
| Fleet map | Shows live balloon positions, selection, trajectories, and operational context |
| Balloon detail | Displays detailed telemetry and per-balloon drill-down from the map |
| Ontology viewer | Displays ontology entities, properties, and relationships as a graph |
| Agent sidebar | Answers fleet and weather questions, focuses the map, and summarizes operational state |
| Security | Uses Entra JWT validation, Managed Identity, rate limiting, and sanitized backend errors |
| Path | Purpose |
|---|---|
specs/001-* to specs/005-* |
Feature specs, quickstarts, and per-feature installation or deployment guides for the simulator, data platform, alerts, ontology, and MCP layer |
src/Workload/Frontend |
React workload UI: map, ontology, agent sidebar, Fabric bootstrap |
src/Workload/Backend |
Express API: chat, telemetry, ontology proxy, auth, health, webhooks |
src/mcp-servers |
MCP servers kept in the repo, with skynav-wind now actively bridged by the backend for external weather and wind |
specs/006-skynav-agent |
Workload specification, architecture, deployment documentation, contracts |
docs |
Supporting documentation and integration notes |
The Fabric workload was initially developed on branch 006-SkyNav-Agent and is now documented here on main together with the supporting platform features.
- Frontend hosting: Azure Static Web App
- Backend hosting: Azure App Service
- Model hosting: Microsoft AI Foundry, using a
gpt-5.2deployment exposed through the Azure OpenAI-compatible endpoint - User authentication: Fabric workload token validated by the backend through Entra JWKS
- Backend access to Fabric and model resources: Managed Identity
- Production path: no client secrets required for the main runtime
Real-time 3D visualization of the balloon fleet using CesiumJS on a globe with live telemetry via WebSocket. The viewer embeds the IoT telemetry simulator — no separate process needed. Balloons are rendered as montgolfière icons with color-coded flight phases, trails, and vertical tethers to ground.
- Python 3.11+ installed and available in
PATH - Internet access (CesiumJS is loaded from CDN, OpenStreetMap tiles as fallback)
- Optional: an Azure Event Hubs namespace for dual-publish (telemetry goes to both the browser and Event Hubs simultaneously)
- Optional: a Cesium Ion access token for high-definition 3D terrain (without it, the viewer uses OpenStreetMap tiles)
# Windows — launch with 20 balloons, opens browser automatically
.\launch.ps1 -FleetSize 20
# Linux / macOS
chmod +x launch.sh && ./launch.sh --fleet-size 20The launcher will:
- Create a Python virtual environment (
.venv) if it doesn't exist - Install all pip dependencies from
src/requirements.txt - Load environment variables from
.envat the project root - Start the viewer server on http://localhost:8080
- Open your default browser automatically
- Wait for you to click "Start Simulation"
Copy .env.template to .env and fill in values. Never commit .env — it is in .gitignore.
| Variable | Default | Required | Description |
|---|---|---|---|
SKYNAV_FLEET_SIZE |
from fleet.yaml (3) |
No | Number of balloons to simulate (1–20+). Beyond 3, extra balloons are auto-generated with French names, random pilots, and Loire Valley launch sites. |
SKYNAV_SCENARIO |
normal |
No | Flight scenario: normal, storm, calm. Affects wind layers, propane burn rate, and cruise altitude. |
SKYNAV_EMIT_INTERVAL |
5.0 |
No | Telemetry emission interval in seconds. Lower = faster simulation. |
SKYNAV_VIEWER_PORT |
8080 |
No | HTTP port for the viewer server. |
SKYNAV_EVENTHUB_CONNECTION_STRING |
— | No | Azure Event Hubs connection string. When set, telemetry is published to both WebSocket (browser) and Event Hubs simultaneously. |
SKYNAV_EVENTHUB_NAMESPACE |
— | No | Event Hub FQDN for Entra ID (managed identity) auth. Alternative to connection string. |
SKYNAV_EVENTHUB_NAME |
— | No | Event Hub entity name. Required if using Event Hubs. |
CESIUM_ION_TOKEN |
— | No | Cesium Ion access token for HD 3D terrain with satellite imagery. Without it, the viewer falls back to OpenStreetMap 2D tiles — still fully functional. |
| PowerShell | Bash | Description |
|---|---|---|
-FleetSize 20 |
--fleet-size 20 |
Number of balloons (overrides .env and fleet.yaml) |
-Scenario storm |
--scenario storm |
Flight scenario |
-Port 9090 |
--port 9090 |
Override viewer HTTP port |
-NoBrowser |
--no-browser |
Don't auto-open browser |
-SkipInstall |
--skip-install |
Skip pip install (use after first run for faster startup) |
# Quick demo — 5 balloons, default scenario
.\launch.ps1 -FleetSize 5
# Full fleet, storm scenario, custom port
.\launch.ps1 -FleetSize 20 -Scenario storm -Port 9090
# Fast restart (skip pip install)
.\launch.ps1 -FleetSize 20 -SkipInstall
# Headless mode (no browser auto-open)
.\launch.ps1 -FleetSize 20 -NoBrowser# Linux/macOS equivalents
./launch.sh --fleet-size 20 --scenario storm --port 9090
./launch.sh --fleet-size 20 --skip-install --no-browserThe base fleet is defined in src/config/fleet.yaml (3 balloons by default). When SKYNAV_FLEET_SIZE exceeds the YAML count, extra balloons are auto-generated with:
- French balloon names (Château Volant, Vent d'Ouest, Ciel de Loire...)
- Random Loire Valley launch sites (Amboise, Chenonceau, Chaumont-sur-Loire...)
- Varied envelope types and pilot names
- Staggered launch delays (each subsequent balloon launches later)
| Phase | Color | Description |
|---|---|---|
| Preflight | Gray | Balloon on ground, waiting for launch delay |
| Ascending | Blue | Climbing to cruise altitude |
| Cruising | Green | At target altitude, drifting with wind |
| Descending | Orange | Returning to ground (propane low or cruise complete) |
| Landed | Red | On ground, flight complete |
- Montgolfière icons — Balloons rendered as hot air balloon icons with envelope, basket, and flame, color-coded by phase
- Real-time telemetry — Position, altitude, speed, heading, wind, temperature, propane level updated every 5s via WebSocket
- Flight trails — Colored polyline trails showing each balloon's flight path
- Vertical tethers — Dashed lines from balloon to ground showing height visually
- Dashboard HUD — Active/ascending/landed counts, simulation tick counter
- Info popup — Click a balloon for detailed telemetry table
- Alert simulator — Trigger low propane, rapid descent, or strong wind alerts on any active balloon
- Auto-reconnect — WebSocket reconnects automatically if connection drops
Browser (CesiumJS) <──WebSocket──> Python Server (aiohttp)
│
Simulator Engine
(tick every 5s)
│
┌─────┼─────┐
▼ ▼
WebSocket Event Hubs
(browser) (optional)
The server runs entirely in one process: HTTP server (serves index.html) + WebSocket server (pushes telemetry) + simulator engine (ticks balloons). No external databases required for the viewer.
| Document | Description |
|---|---|
| Global project plan | Broader project plan across simulator, data platform, and workload |
| Toolchain plan | How Agent Forge, Spec-Kit, and Squad are used across the project |
| Feature squads plan | Breakdown of the work into feature squads and contracts-based parallelism |
| Feature 001 deployment guide | Install and run the IoT telemetry simulator locally or with Azure Event Hubs |
| Feature 002 deployment guide | Provision Event Hub, Eventhouse schema, Eventstream, and KQL validation flow |
| Feature 003 deployment guide | Configure Activator rules, schema extensions, alert destinations, and validation |
| Feature 004 deployment guide | Seed reference data, deploy the Fabric ontology, and validate the graph bindings |
| Workload deployment guide | Manual deployment flow for Azure resources, Entra ID, workload packaging, frontend and backend |
| Workload architecture | Detailed runtime architecture, data flows, security model, and deployment diagrams |
| Workload specification | Functional and technical specification for the Fabric workload |
| Deployment via PowerShell | Automated deployment path with deploy-skynav.ps1 |
| MCP servers README | Repository MCP server architecture and development notes |
- The KQL Eventhouse and Fabric ontology are existing Fabric workspace resources consumed by the workload
- The current headless sidebar is the main production chat UX; CopilotKit runtime remains available for compatibility and future streaming scenarios
skynav-windis now used additively from the backend for external Open-Meteo weather and wind lookups- The root README is intentionally concise; use the workload architecture document for the full technical breakdown