conversation with agent
grafana dashboard
trial website
conversation with agent
conversation with agent

Inspiration

In a process plant, the person who first sees an alarm is often walking the plant, not watching the control screen, and one problem lights up a dozen alarms at once. Operators already carry a phone and open WhatsApp without thinking. We wanted the plant to talk to them there: one clear message per problem, an assistant that reads the same data as the control room, and an ERP that hears about the problem from the plant floor, not hours later.

What it does

NeuraPlant Sentinel watches the plant every minute and writes first when something goes wrong:

The plant raises the alarm. Fixed rules, with no AI, check every variable against its operating limits.
One notice per problem. Alerts are grouped into episodes: one WhatsApp notice when an episode opens and one more if it turns critical. Each cycle of the simulated plant, with five incidents, sends 2 notices.
The diagnosis is ready before the operator asks. A Claude agent prepares it in the background: what is happening, which signal moved first, whether both temperature probes agree and which maintenance orders are open.
A conversation alongside the dashboards. The operator asks follow-up questions in plain language and checks the answers on the Amazon Managed Grafana plant dashboard.
From the plant floor to the ERP. When the operator confirms, the agent logs the incident and raises the SAP PM maintenance notification, and the operator gets its number on WhatsApp.

How we built it

Plant to cloud: an Ignition simulator plays a looping 20-minute program (15 minutes of incidents, 5 of normal operation); its tags travel over OPC UA to a SiteWise Edge gateway and into AWS IoT SiteWise.
Alert pipeline: four AWS Lambda functions chained with Amazon EventBridge (evaluate limits, register alerts and episodes, notify on WhatsApp, prepare the diagnosis), with their state in Amazon DynamoDB.
Agent: Claude Sonnet 4.6 on Amazon Bedrock, deployed on Amazon Bedrock AgentCore Runtime with one session per operator. Its seven tools are served through AgentCore Gateway (MCP, IAM only) by a tools Lambda, and the agent reads the tool catalog from the Gateway.
ERP: SAP PM orders and notifications simulated in DynamoDB behind one module, ready to swap for the SAP S/4HANA APIs.
Dashboards: Amazon Managed Grafana shows the SiteWise signals with their limit bands, plus episodes, WhatsApp notices and SAP notifications from Amazon CloudWatch Logs.
Delivery: the infrastructure is created with Terraform through a GitLab pipeline (only the WhatsApp number and the seed data are set up by hand). Offline tests replay the simulator, and an LLM-judged evaluation scores the agent on five plant scenarios. -Pre-existing and third-party code: everything in the repository was written during the hackathon, except three generic Terraform stacks the team already had (network, ec2 and the NeuraPlant data platform), which are copied in iac/stacks/. Ignition, AWS IoT SiteWise Edge, Amazon Bedrock with Claude and Meta's WhatsApp Business platform are used under their own terms.

Challenges we ran into

Alarm floods: a thermal runaway moves several signals at once, and our first version sent 10-12 messages per cycle. Grouping alerts into episodes cut it to 2, without process-specific rules.
Late, batched plant data: SiteWise Edge uploads readings about a minute late and in batches, and a short pump trip slipped through. The evaluator now judges every reading in a window.
WhatsApp rules: a business can only write first with a template, so the notice goes out at once and the diagnosis is prepared in the background, ready when the operator taps the button.

What we learned

Keep the AI where judgement is needed: detection and notification are plain code; the agent explains, correlates and writes to the ERP only when the operator confirms.
Messaging rules shape the design as much as any technical choice.
Industrial real time means a minute or two: the logic has to judge windows of readings, not single values.

What's next

Plant procedures in an Amazon Bedrock Knowledge Base, AgentCore Memory to learn from past incidents, the real SAP S/4HANA APIs, shift handover and escalation, and SMS or RCS where WhatsApp is not allowed.