Campus Digital Twin
Unity · C# · Java · TCP · MQTT · CSV · NVIDIA Omniverse · USD · FastMCP · AWS
- Twins shipped
- 3
- Live data sources
- HITS data lake, MQTT, CSV
- Ownership
- Architecture, connectors, deployment
Problem
A campus is a system of systems: network labs, energy infrastructure, and operational spaces that behave differently under the same conditions. The university had data — telemetry, logs, plant readings — but no way to see how a change in one part of the campus would propagate through another. Static dashboards showed what had happened. They could not answer “what if”.
Constraint
The twins had to be driven by the live HITS data lake, not by curated demo files. That ruled out a offline authoring tool: the simulation layer had to ingest, align, and react to real streams. It also had to be built by one engineer and handed over to researchers who are not engine specialists.
Decision
I built three twins on one shared architecture rather than three separate demos:
- Network lab twin — Unity simulation backed by Cisco Packet Tracer through a custom Java/TCP bridge. Topology state flows both ways, so lab exercises can be rehearsed against a live-equivalent device layer.
- Photovoltaic plant twin — a physics-based model of a live solar installation, connected to plant telemetry.
- First-person cybersecurity puzzle game — an interactive training environment built on the same simulation spine.
The connectors are the reusable part. A CSV/MQTT ingest extension reads scheduled or streamed data and writes it into USD attributes and OmniGraph, so an Omniverse scene reflects the current state of the physical plant. The same data lake is exposed to an AI agent through a FastMCP server — stdio locally, Streamable HTTP on EC2/Fargate and Lambda — including a tool that returns live PV inverter readings.
Verification
Validation was structural, not decorative:
- Each twin was checked against the live source it was bound to, not against a scripted scenario.
- The ingest path was exercised with real data lake connections, then with replayed CSV for regression.
- The MCP server was tested over both transports (local stdio, deployed HTTP) to confirm the agent sees the same tool surface.
- Deployment was part of the work: the connectors and agent endpoint run on the university’s infrastructure, not on a developer laptop.
Result
Three twins shipped and connected to live campus data. I owned the architecture, the data lake connectors, and deployment end to end — the same module from requirements to running system.
The transferable lesson: a digital twin is only as good as the fidelity of its link to reality. Most of the engineering was in the link, not in the visuals.