The $25,000 Search Box That Taught the Pentagon to Buy Software

The first product Vannevar Labs tried to build was smart, difficult and nearly beside the point. The young company spent months on Arabic optical character recognition, the business of persuading computers to read text that makes ordinary scanners perspire. There was only one defect: the problem was not urgent enough. In the Pentagon, where procurement can move with the stately confidence of a glacier, a merely interesting problem is another way of saying no.

The change of mind came not from a strategy retreat but from watching demonstrations. Vannevar tried a small intelligence-search prototype called Decrypt: one primary feature, one modest data source, two engineers. People interrupted the demo. They wanted it on deployment. Soon, four users bought a three-month pilot for roughly $25,000 and began producing intelligence that travelled upward through government. The prototype was limited. Its usefulness was not.

The useful failureA product should arrive with a pulse

That early reversal is the most important fact about Vannevar, more revealing than a funding round and more useful than any mission statement. The company learned that defense products do not win because their technology is broadly impressive. They win because a person with a live mission would be inconvenienced if the product disappeared tomorrow.

Decrypt searches, translates and organizes foreign-language and open-source material for analysts, operators and policy teams. It was launched publicly in 2021 and grew from one deployment to about a dozen adoptions that year. Its descendants widened the aperture. Serra collects and analyzes non-traditional and geospatial intelligence. Newer mission-specific agents model adversary behavior, handle analytical chores and help users move from a sea of signals to a smaller set of choices. Vannevar also integrates software with sensing systems for tactical environments where connectivity is unreliable and a public-cloud demo is as useful as a silk umbrella in a hurricane.

“That kind of relationship you only build by being really useful to start.”Brett Granberg, co-founder and CEO

The customer is not a logoFour users can be a market if the mission matters

Vannevar's customers are government organizations, but its users are people with narrower nouns: analyst, planner, operator, commander. They work across service branches and combatant commands, often with multilingual information, open-source data and sensor feeds arriving faster than a human team can sort them. The problem is not a lack of data. It is the comedy of having too much data and too little time.

The company's business model follows the peculiar staircase of federal buying. Start with a paid pilot or prototype. Prove value in an operational setting. Add users, data and mission areas. Then seek a production contract or multi-year program. In 2024 the Defense Innovation Unit moved Vannevar's prototype work into a production agreement with a potential ceiling of $99 million, of which $16 million was obligated at award. A ceiling is not revenue, a distinction that enthusiastic headlines often misplace. It is permission to earn revenue up to that amount.

A moat made of proximityThe engineer goes to the problem

The mechanism is forward-deployed engineering. Engineers embed with users, learn the mission and feed what they see back into product design. Serra emerged this way: not as a fashionable square on a roadmap, but from an engineer close enough to notice a recurring need. The company pairs that fieldwork with a reported national-security dataset exceeding one petabyte and growing by about ten terabytes a week. Models trained on generic internet text can speak fluently about many things; Vannevar's wager is that military usefulness depends on knowing the rare, awkward material specific to the adversary and the mission.

This also explains the staff. Co-founder and president Nini Hamrick spent seven years in U.S. intelligence, including time embedded overseas and as chief of staff to the director of the National Counterterrorism Center. CEO Brett Granberg came through national-security technology investing. Their company deliberately combines operators with software engineers, machine-learning researchers and people from quantitative finance. The culture manual uses the phrase “team of Jedi,” which is not a sentence every adult can deliver safely. The more serious principles are proximity, accountability, low ego and sustainable intensity.

The crowded mapNot Palantir, not Anduril - and not entirely unlike either

Vannevar occupies the lively territory between intelligence software and the new defense primes. Palantir is an alternative for operational data platforms. Primer and Babel Street compete in AI-assisted intelligence analysis. Anduril combines software and hardware at a far larger scale. Legacy contractors and home-grown spreadsheets remain formidable rivals because procurement habit is the most deeply installed software in Washington.

Vannevar's distinction is its concentration on non-kinetic missions - intelligence, information operations, influence, maritime awareness and decision advantage before weapons are fired. Its partnership with SMX gives it access to larger contract vehicles and delivery infrastructure. Work with Ultra shows its agents enriching battle tracks inside another company's platform. This is less a castle than a network: proprietary data in the center, mission teams at the edge and partners carrying the product through doors a young contractor cannot open alone.

What travelsCopy the sequence, not the camouflage

The Vannevar lesson works outside defense. First, rank problems by urgency rather than intellectual charm. Second, make the first version almost embarrassingly narrow. Third, charge early; payment reveals seriousness that compliments conceal. Fourth, put builders beside users. Finally, let a measurable win create the next sale. The sequence is obvious in print and rare in organizations, which prefer to debate platforms because platforms postpone the vulgar moment when somebody must use the thing.

There are conditions attached. The method depends on access to real operators, patient capital and a buyer able to scale a successful pilot. It is expensive when engineers must travel, meet security requirements and support disconnected environments. It also falters when outcomes cannot be discussed, measured or carried by an internal champion. A bespoke solution can become a consulting project wearing software's margins; a narrow pilot can remain narrow forever. Vannevar's production transition matters precisely because most defense prototypes never enjoy one.

What began as a search box is now presented as an agentic AI platform supporting more than 125 deployments. The company has raised more than $90 million, yet the original lesson remains delightfully small. A product does not need to look inevitable. It needs four people who would object if you took it away.