25 March 2026 – By Arthur Karell
As the cost of deploying AI at the edge collapses, competitive manufacturers and integrators are going to want their own autonomy stack in their products, even down to the component level. An unmanned vehicle will have its own onboard autonomy software, its alt-GPS subsystem will have its own autonomy stack for navigation, various weapon systems will have their own autonomy software for targeting/fire control, and the integrated system will be communicating with third party command and control AIs. An OEM without proprietary autonomy in its products will be rapidly commoditized, and worse, won’t have the capabilities to meet mission. How will these OEMs and system integrators build these autonomy stacks at pace and cost?
Enter Vanguard Defense. Johnny Ni and his team have built a vast repository of domain-specific datasets that allow OEMs and other vendors to take COTS AI models and create autonomy stacks that are performant and useful for their end customers. Brilliantly, they began by partnering with the Veterans Administration to create a network of cleared veterans who find it very meaningful to put their domain expertise to work laying the foundation for Vanguard’s data observability products.
It’s the kind of double-bottom line win that we love and the demand has been tremendous, because models and algos are increasingly not the limfac for new AI products. The real constraint is often the data.
Security-specific datasets differ significantly from the types of datasets that powered the first wave of commercial AI. They are smaller, noisier, and far more dependent on operational context. A radio signature in an electronic warfare environment or an image captured by a drone over a complex maritime facility is not an interchangeable data point. Its value depends on conditions, equipment, and behavior that someone with domain knowledge must recognize.
That requirement creates a bottleneck.
Most of the data-labeling infrastructure built during the commercial AI boom relied on generalized labor pools performing standardized tasks at scale. That approach works well for consumer datasets. It does no work in sensitive environments or when mission-critical accuracy depends on expertise. As AI expands into sectors like defense, aerospace, and advanced manufacturing, the value of expert labeling increases dramatically.
Vanguard was built around that premise. It’s more than a workforce; that network sits within a broader stack that includes workflow infrastructure, quality benchmarks, evaluation frameworks, and proprietary tools making expert labeling repeatable at scale.
At First In, we spend a great deal of time identifying companies positioned at structural bottlenecks. These are the parts of the system that quietly become indispensable as an industry grows, and Vanguard is building essential long-term value at the AI infrastructure layer. We are proud to lead their Series Seed with our first investment from First In’s SBIC Critical Technologies Fund 3.