.png){kind=logo}
At Initium Technologies, Scheibye directs an engineering team building edge AI systems for military environments where limitless compute is not available.
An AI model can perform well in testing and still be unsuitable for a deployed military system if the surrounding hardware, software stack, and operating environment cannot support it. Conrad Scheibye deals with that problem at the level of the complete system. As co-founder and CTO of Initium Technologies, he directs the technical work behind edge AI systems designed to process sensor data close to where it is collected, including environments where reliable access to remote computing infrastructure cannot be assumed.
Scheibye said,“Deploying AI at the edge introduces constraints you don't hit anywhere else, and they are opposite to what you experience when using cloud compute. Everyone sizes the problem around the accelerator. For us the accelerator is rarely what runs out first. It's the video decode path, or memory bandwidth. The GPU sits half idle while something else is what's actually capping you. None of this shows up in the cloud, where vCPUs and bandwidth are plentiful and the GPU is the only thing you're really rationing.”
Commercial AI infrastructure often assumes access to centralized computing resources and a dependable network connection. Military operations cannot rely on those conditions. Connections can be slow or intermittent, electronic warfare can disrupt communications, large sensor streams can rapidly consume available bandwidth, and operational detachments may deliberately limit reachback while maintaining stealth. Those conditions shape the architecture from the beginning.
According to Scheibye, “In active theaters DDIL conditions and self-imposed reachback limitations are not uncommon. This prevents operators from using the cloud to run software capabilities. Edge compute systems allow them to leverage these otherwise unavailable capabilities to ensure mission success”
Thermal design is one example of where those tradeoffs become physical. Initium evaluates enclosures that have to keep high-power computing modules inside acceptable temperature ranges in both desert and tundra conditions. More compute can increase cooling requirements, which can affect the physical characteristics of equipment intended to travel with a unit or mount on a platform. The technical question becomes whether the complete system can deliver useful capability inside the intended operating environment.
IRIS provides one example of that system's problem. The backpack-mounted platform is designed to perform computer-vision tasks locally across existing camera feeds rather than depending on remote processing. IRIS poses a concrete systems problem: model performance, hardware capability, thermal behavior, and portability all have to resolve into equipment that can function in the intended operating environment.
“Being an operator-carried system, IRIS not only needs to function in its intended environment, but must be sufficiently ruggedized to survive the trip. For example, if a team inserts via a subsurface infiltration, IRIS needs to withstand the pressures felt at 30 FSW, despite never being used underwater,” Scheibye said.
MANTIS creates a different architectural decision. The system is intended for aerial and maritime surveillance, where continuous sensor feeds can consume a large share of a tactical connection. Its architecture moves processing closer to the sensor so that detections and geographic positions can be transmitted instead of requiring all raw data to compete for limited communications capacity.
That architecture changes what has to cross the tactical connection. MANTIS is designed to perform more processing near the sensor so detections and geographic positions can travel instead of the complete raw feed.
MANTIS was created in the early stages of Initium, initially Conrad developed the model to detect objects from drone feeds, the geolocation capabilities as well as the STANAG 4609 encoder and decoder on his own.
Initium has traveled to military installations around the United States to demonstrate systems and gather feedback from prospective users. Those interactions feed back into technical planning. Requirements from the field can expose assumptions in the architecture that do not survive contact with the intended workflow, and technical leadership has to determine what should change in response.
That is where Scheibye’s contribution differs from ownership of a single subsystem. He has to determine whether model performance, hardware capability, thermal behavior, communications assumptions, and operator requirements are converging on a deployable system. At the CTO level, Conrad has to decide whether the team’s specialized work has converged far enough that the complete system can meet the operating requirements being set by its intended military users.
