Maintenance Agent

Inspiration

Industrial plants generate continuous machine telemetry, but maintenance teams still spend most of their time manually investigating alerts: checking history and manuals, guessing severity, writing work orders, and chasing technicians. Alerts create noise; investigation and action stay slow.

We wanted an agent that does not wait for a chatbot prompt — one that wakes up when a machine fails, investigates across systems, and takes real maintenance actions.

What it does

Maintenance Agent is an event-driven autonomous workflow (Taskmaster track):

1. Simulated fleet telemetry streams in (temperature, vibration, motor current)
2. Code-based anomaly detection opens an incident when limits are exceeded
3. A Gemini 3.5 agent (via Google ADK) investigates live using tools
4. The agent diagnoses the failure, checks inventory, creates a work order, notifies a technician, and updates machine status
5. The React dashboard shows the investigation unfolding step by step
6. After repair, Mark as completed triggers verification — the agent checks post-repair telemetry and resolves the incident if healthy

Important: telemetry can be simulated; the investigation, tool calls, and decisions run for real.

For CRITICAL shutdown recommendations, a human must Approve or Reject — the agent does not shut down machines on its own.

How we built it

• Model: Gemini 3.5 Flash on Vertex AI
• Agent framework: Google Agent Development Kit (ADK)
• Backend: FastAPI on Google Cloud Run
• Data / events: Cloud Firestore and Pub/Sub
• Frontend: React (Vite) operations dashboard

Tools cover machine context, telemetry history, maintenance records, manuals, inventory, work orders, notifications, and approvals.

telemetry → anomaly detection → incident → ADK agent investigates
  → diagnose → inventory → work order → notify → update status
  → (after repair) verify → resolve

Challenges we ran into

• Making autonomy visible: returning only a final summary hid the real work. We exposed mid-run tool progress in a live investigation timeline.
• Keeping the demo honest: no prewritten Gemini diagnosis for the main path — simulated sensors, live agent.
• Simulator timing: healthy period → degradation → limit breach had to be clear enough for a short video without dead air.
• Responsible autonomy: auto-shutdown needed a human gate without blocking the main work-order path.

Accomplishments that we're proud of

• A full closed loop: detect → investigate → act → verify → resolve
• A dashboard where judges can watch real tool calls appear during a live run
• Real ADK/Gemini actions that write work orders and status into the same store the UI reads
• Deployed on Google Cloud (Cloud Run, Firestore, Pub/Sub, Vertex AI)

What we learned

• Framing the product as an event-driven agent (not a chat UI) makes autonomy easier to judge: tools and state changes are the deliverable.
• UI observability matters as much as the model — a live activity timeline sells the agent better than a final paragraph alone.
• ADK kept instructions and tools in one place; the backend triggers the agent and persists results.
• Human approval for shutdown is a strong demo of safe autonomy.

What's next for Maintenance Agent

• Richer failure modes and multi-agent handoffs (diagnostics vs scheduling)
• Real CMMS / plant data integrations beyond seeded demo data
• Stronger streaming of agent reasoning into the UI
• Production hardening: auth, audit trails, and stronger safety guardrails

Built With

• fastapi
• firestore
• gemini
• google-adk
• google-cloud
• google-cloud-run
• pubsub
• python
• react
• typescript
• vertex-ai
• vite