Smart-Plant

An intelligent multi-agent system that monitors factory operations, diagnoses issues, and proposes actions — all with human-in-the-loop approval.

Comment

Inspiration

I wanted to see if AI agents could actually be useful on a factory floor — not just as a chatbot, but as something that watches your machines, catches problems before they escalate, and hands decisions back to the human operator before acting. Most industrial monitoring tools just alert you. We wanted one that could think and act.

What it does

Smart Plant monitors factory sensor data in real time and runs a set of AI agents in the background. If a motor temperature spikes, it flags it and drafts a maintenance email. If the data pipeline goes down, it detects the failure, proposes a fix, and waits for a human to approve before doing anything. Operators see pending actions in a dashboard and can approve, edit, or reject them before anything gets sent or executed.

How we built it

The backend is a set of agents built on Google ADK and Gemini 2.5 Flash, running on Cloud Run. Sensor data flows through Fivetran into BigQuery, where the agents query it. The frontend is a Next.js dashboard deployed on Vercel. Actions queue in BigQuery and the Action Agent polls for approvals every 10 seconds.

Challenges we ran into

BigQuery's streaming buffer blocks UPDATE queries for up to 90 minutes, which broke our approval flow entirely — we had to switch to DML INSERT with parameterized queries. We also kept hitting Vertex AI's rate limits when multiple agents fired at once. Getting the Fivetran auto-repair to actually unpause the connector took more debugging than expected.

Accomplishments that we're proud of

The full HITL loop works end-to-end: sensor anomaly detected → action proposed → operator approves → email sent or pipeline repaired automatically. The pipeline card on the dashboard shows live Fivetran connector status and flips red the moment it goes down. What we learned AI agents in production hit real infrastructure limits fast — quota limits, streaming buffers, API quirks. The hardest part wasn't the AI, it was making the plumbing reliable enough for the agents to be useful.

What's next for Smart Plant

Connecting more sensor types and production lines, adding predictive maintenance (flagging issues before they breach thresholds), and tightening the agent feedback loop so the system learns from past approvals.

Built With

Share this project:

Updates