Digital twin and agentic AI for energy efficiency

The challenge

The data existed, but it was scattered across meters, sensors and technical manuals. Operators only saw where energy was wasted at the end of the month, and finding anything in the documentation took hours. The client wanted AI that is genuinely useful, not a demo chatbot.

How we worked

• Discovery through interviews with technicians and plant managers: the questions they ask every day became the assistant’s backlog.
• Incremental releases every two weeks on a pilot plant, with adoption metrics measured before rolling out to other sites.
• A set of test questions with expected answers, used as an automated check on every change to the AI.
• Operator training and a direct support channel during the first weeks of use.

Architecture and choices

• The real-time data path is separate from long-term storage, so a slow archive never delays the dashboards.
• The digital twin as a reading layer: every sensor is linked to its asset, so an anomaly shows up at the exact point in the plant where it happens.
• An agentic assistant that picks the right tool (time series, documentation or both) through typed functions instead of free-form generated queries.
• Every answer cites the data or document it comes from, so it can be checked.

External services, and why these

Outcome

In use on the pilot plant, with rollout to other sites under way. The client is covered by a confidentiality agreement.

Stack

• AWS IoT Core
• MQTT
• AWS Lambda
• PostgreSQL
• pgvector
• Amazon Bedrock
• Python
• React
• Three.js