Trials in real environments
Pilot studies with operating power plants, running agents on live measurement data to study how reliable their workflows really are.
Trustworthy, autonomous
cyber-physical systems
A research initiative to develop a distributed agentic-AI system for cyber-physical systems.
Research challenges
Agentic AI in critical industrial settings is not a solved engineering problem. Agents that pass benchmarks may still fail on real tasks. Evaluation method, not model size, is now the bottleneck. These are the open problems shaping what we work on.
Accuracy is not enough
Industry needs holistic measures (correctness, safety, cost, compliance), not task-success alone.
Acting agents, real consequences
Agents that act on the physical world need provable guardrails and audit, not aspirational ones.
Inference at the edge
How to build edge servers for industrial agentic AI, and keep utilisation high enough that on-prem inference stays viable.
Interfaces synthesised on demand
Task-specific interfaces generated by the agent and discarded after, replacing static, monolithic apps.
Analyses on the fly
Agents that generate the analyses, dashboards and reports a domain expert would otherwise hand-craft.
Example applications
Conversational agent
Engineers task agents in natural language; every step is traceable.
Long-horizon agents
Define agent personas, tools and routines; deploy them across operations.
Building intelligence
3D building twin the agent navigates, queries, and acts on in the loop.
Wall thickness analysis
Inspection workflows over ultrasonic measurement data, with auditable reports.
Shared operational memory
A single distributed memory every agent reads and writes. What one agent learns, every other agent can act on.
Terminal user interface (TUI)
Developer surface for running, inspecting and debugging agentic workflows.
Example use cases
Testing & inspection
Agents draft inspection reports, cross-check measurements against standards, and hand off to the engineer for sign-off.
Operating the physical world
Agents reason over sensors, actuators and digital twins, and act in the loop in natural language.
Plant operations
Agents watch every signal across operations, surface anomalies early, and remember what has been seen.
Predictive maintenance
Agents project remaining life of bearings, pumps and motors from vibration, temperature and load history.
Grid & generation
Agents forecast demand, balance generation, and triage faults across distributed assets in real time.
Production line AI
Agents close the loop on quality, defect detection and yield, batch by batch, with full provenance.