A company that sells the opposite of a black box
Here is a thing that is true about enterprise artificial intelligence, and that everyone in the industry politely agrees not to dwell on: the scariest part is not that the model might be wrong. It is that when the model is right, nobody in the building can fully explain why. You feed a vendor's system your most sensitive text - trades, clinical notes, intelligence reports - and it hands back an answer with the confidence of a weather forecast and the auditability of a horoscope. You are asked to trust it. So you do, a little nervously, and you hope the auditors don't ask.
dMetrics, an AI company founded in 2011 by two MIT machine-learning PhDs, built its entire product around not doing that. Its platform, called Minsky, lets non-technical subject-matter experts - analysts, researchers, the people who actually know the domain - build and control their own AI models for reading text. No coding required. The pitch is almost aggressively unglamorous: the expert builds the model, so the expert trusts the output. Transparency is not a compliance checkbox bolted on at the end; it is the reason the thing exists.
If you want to understand a company, it usually helps to know what it was doing before it was doing the thing it is famous for now. dMetrics is famous now for defense work. But it started, improbably, by reading what people said about their medications on Facebook.
From drugstore chatter to knowledge graphs
In the late 2000s at MIT, Paul Nemirovsky - a Media Lab graduate student, PhD '06 - teamed up with Ariadna Quattoni, a CSAIL researcher who earned her computer-science PhD in 2009. Their opening ambition was the kind of sentence that gets you either funded or laughed out of the room: use big data to make everyone an expert. They noticed that healthcare was the thing ordinary people worried about most, and that the internet was drowning in unstructured chatter about it. So they built DecisionEngine, a platform that read billions of online conversations about drugs, devices, and health products across blogs, forums, Facebook, and Twitter, and tried to extract something useful from the noise.
The technical insight is the part worth keeping. Most tools of that era ran on sentiment analysis - thumbs up, thumbs down, a happiness score. Nemirovsky's objection was that humans do not speak in scores. "Language and expression doesn't work like that," he told MIT News. "We're a bit more complex as humans." DecisionEngine used a layered funnel instead: filter the noise, separate genuine personal experience from marketing, and only then identify the actual decision a person made. It was, in retrospect, a very early bet that reading text well was harder and more valuable than measuring mood.
The company grew from two people to sixteen, collected four National Science Foundation grants, moved from Boston to Brooklyn, and quietly picked up Fortune 500 clients in healthcare, then finance. Somewhere in that decade the product matured from "analyze health conversations" into something more general and more powerful: a platform that could build natural-language-processing models and knowledge graphs out of sparse, messy data on almost any subject. They named it Minsky, after Marvin Minsky, one of the founding figures of AI - a confident name, the kind you have to then spend years earning.
The part where the Pentagon calls
In 2019, the Defense Innovation Unit - the arm of the Department of Defense whose whole job is to find commercial technology and drag it into government use - teamed up with the U.S. Army and the Defense Technical Information Center on a specific problem. They wanted a machine-learning tool that could collect, analyze, and report on threat activity buried in web-based, open-source content. The internet says a great many things; most of it is irrelevant; some of it is not; sorting the two at scale is exactly the messy-text problem dMetrics had spent a decade on.
In March 2020, dMetrics was selected out of a pool of 65 companies to build Minsky as a prototype for an effort called AI-Based Knowledge Graphing. What made the difference, by the government's own account, was not a flashy demo. It was that the customer's own analysts could operate the tool. Minsky lets an analyst spin up a personalized ML agent that continuously scans large publicly available and commercially available datasets on their behalf, then identifies and extracts the entities, actions, and flashpoints relevant to that analyst's specific area of responsibility. The analyst does not file a ticket with a data-science team and wait three weeks. They build the thing themselves.
That prototype worked well enough to graduate. The Department of Defense awarded dMetrics a five-year production contract with a ceiling of $99.5 million, structured as a Production Other Transaction agreement designed so that multiple organizations - Washington Headquarters Service, the Defense Technical Information Center, the Army Geospatial Center, and various intelligence-community agencies - can all deploy the platform. Nemirovsky, now signing statements as "Dr. Paul Nemirovsky, CEO and Cofounder," called it what it was: a bet by the DoD on cutting-edge AI to keep a technological edge.
The financial oddity worth noticing
Now here is the detail that a certain kind of finance-brained reader will enjoy. dMetrics did not get to a nine-figure government contract by raising a nine-figure venture round. Its disclosed equity funding is small - an angel round of roughly $2.3 million, with total disclosed funding in the neighborhood of $2.48 million, plus those early NSF grants. Third-party estimates put annual revenue around $1.2 million before the production contract. The company has about 28 employees. It is headquartered, of all places, in Jackson, Wyoming.
So you have a small, capital-efficient, geographically improbable company that spent roughly a decade being patient and technically stubborn, and then landed a contract worth many multiples of everything it had ever raised. This is not the standard Silicon Valley story of burning $200 million to buy growth. It is closer to the opposite: stay small, stay sharp, wait for a customer whose problem is exactly your problem, and let the tool sell itself in the room. Whether that is a repeatable strategy or a happy accident is the kind of question that keeps founders up at night. But it happened.
The through-line, from healthcare chatter to defense knowledge graphs, is remarkably consistent. dMetrics has always believed that the valuable work is reading unstructured text precisely, and that the person best equipped to judge whether the reading is correct is the domain expert, not the model's authors. In a moment when much of the AI industry is racing to remove humans from the loop, dMetrics is selling the loop itself - a way to keep the expert in the driver's seat while the machine does the reading. It is a quieter thesis than most. It also, so far, has a government check attached to it.