Groundwork is an AI maintenance engineer that lives on the plant floor. It watches rotating equipment on its own and flags a developing fault before the line goes down.
When the hop-level sensor data changes, it diagnoses the fault, reads the equipment manuals and the CMMS history, and opens a work order with citations. A technician does not have to ask first.
NemoClaw runs the whole agent on one Dell Pro Max GB10 with Qwen3.6-35B-A3B. Inference stays on the box. There are no cloud calls. OpenShell sandboxes the process and records a deny if anything tries to send telemetry out.
The technician board is a live plant view: what needs attention, why it happened, and the next action.
Built for the Dell × NVIDIA AI Hackathon at Cornell, 12 Sep 2026.
Public demo: get-groundwork.vercel.app. Pushes to main deploy it through GitHub Actions. groundwork.vercel.app and try-groundwork.vercel.app are already taken on other Vercel accounts.
How it works
The always-on hop loop writes L1 historian tags (not raw 12 kHz waveforms) into SQLite. Diagnosis goes through predictive-maintenance-mcp. Groundwork does not rebuild that DSP. Every claim on screen points at a real artifact: a CWRU or Mendeley window, an SKF manual page, or a CMMS row.
If you tell it to email or POST plant data off-box, OpenShell's L7 policy denies the call and the gateway audit log is what the board shows. Containment is not a seeded string.
Run on the GB10
git clone https://github.com/mdelriolanse/groundwork.git
cd groundwork
node app/server.mjs
Landing: http://127.0.0.1:8765/
Board: http://127.0.0.1:8765/prototype/
Hop loop (separate process):
python3 scripts/hop-loop.py
Inject a catalogued fault for the demo:
./demo/inject/inject-ir.sh
Local inference is http://127.0.0.1:8000 serving Qwen3.6-35B-A3B. Copy weights to internal NVMe before serving. Do not serve off the HACKPACK USB.
npm test
Layout
app/board API, assist, OpenShell evidenceweb/prototype/technician workspaceweb/twin/3D plantscripts/hop-loop.py1 Hz watcher
Log in or sign up for Devpost to join the conversation.