Generative AI for Operational Technology

The DoD relies on a complex mix of legacy and modern critical infrastructure, or Operational Technology (OT for short), to maintain day-to-day operations and readiness. These systems number in the thousands -- think airfield lighting, security alarm, fire alarm, building automation, heating/cooling, generators, water/sewer, and test equipment.áDecision makers need data from OT to make timely, well-informed decisions, and oftentimes lack critical information that is locked into organizational, political, or technical siloes. SimplesenseÆs Installation Resilience Platform (IRP), now deployed across Tyndall AFB through the Installation Resilience Operations Command & Control (IROC) project, has been a huge success, unlocking the siloed data from hundreds of buildings and tens of OT systems.áThe problem now is -- there is too much information for any single human to consume at any one time. In fact, IROC at Tyndall generates over 120GB of data per week in network logs, OT system logs, sensor data, and cybersecurity data.áFor this effort, Simplesense proposes applying a Large Language Model to the 120GB of data generated weekly at Tyndall AFB and working with warfighters to understand the questions they ask of that data and tune the model to effectively answer those questions.áCurrently, during an incident, operators and decision makers pick up the phone and call multiple people to get information. By one operatorÆs estimate, each incident requires at least 20 minutes of data gathering, with multiple incidents a week. An interactive LLM could save hours a week operationally and critical minutes during a severe incident when every second counts.áFrom a mission assurance perspective, a properly tuned LLM can tie together disparate data sources much faster and in ways that a human wouldnÆt think to. For example, an LLM could suggest that a cybersecurity alert is linked to equipment failure faster than a human could, putting cyber defenders days ahead.