Oluwadi filed us under a regional edition instead of a national AI list. That is the right unit of analysis, and it is worth saying why.
Oluwadi recently wrote up our paraphrase-chain study and TeraContext.AI’s classification pipeline under a title I could not have improved on: A Venture Studio Ran 17 AI Models Through a 30-Round Game of Telephone and Published the Results. One of Its Portfolio Companies Builds Construction Bids on the Same Discipline. For anyone new here: Joshua8.AI is the venture studio, and TeraContext.AI is the first product company it is incubating. Oluwadi’s read was that the studio and the company are running the same discipline at two altitudes: one testing models in public, the other testing outputs in production for a customer whose margin depends on the gap between “close enough” and “exactly right.” I will take that.
This note is not about the study. It is about the shelf they put it on.
Oluwadi did not drop Joshua8.AI into a national heap of AI startups. It put us in a regional edition whose header is not “greater Washington tech scene.” It is DMV — Defense · Intelligence · Technology Capital. And the DMV edition is not a courtesy to a local company. It is the edition Oluwadi launched with in July, and the site is built region-first all the way down. There are 108 regional editions. The four entry points come in a fixed order: Region, then What a company builds, then Who it serves, then Analysis. All of it sits on the same graph of companies, investors, and people that powers Osparna, the paid diligence instrument behind it. Oluwadi is the free front door, and region is the first question it asks, before sector, before stage, before who raised.
That ordering is the whole point, and it is the opposite of how most directories work.
A Bay Area list is a gravity well
It sorts companies by who raised, who sits near a famous lab, and who resembles the last company that looked like this one. That is a reasonable index for Sand Hill Road. It is a bad index for this metro. Oluwadi’s own analysis of the region, DMV: Not More VC, a Different Kind of Capital, puts numbers on it: investors and VC firms are 52 percent of the companies in their Silicon Valley data and 24 percent here, defense and aerospace run about four to one in our favor, and a meaningful slice of the capital in this region moves through a channel where the customer is a government and the return is a contract or a mission outcome. I would only add that the same thing is true one layer down, where the customer is a general contractor and the return is a bid that did not miss a section.
I have seen the gravity well from inside it. Around 2000 I started an optical networking company in Silicon Valley, funded by two Sand Hill Road firms, Benchmark Capital and U.S. Venture Partners. Both partners who led that round said the same thing in the first meeting: they intended to be in the building, in person, and they would not drive more than thirty minutes to do it. That was not laziness. It was how they added value, and they knew it. But it also meant that the fundable universe, as seen from Sand Hill Road, was a thirty-minute radius. Everything outside the circle was somebody else’s problem, or nobody’s. Regional investing is not a sentiment. It is a driving distance, and a directory that starts with region is starting where the capital actually starts.
One metro, several industries that do not share a narrative
The overlooked part is not that Northern Virginia has companies. Anyone who has landed at Dulles knows that. The overlooked part is that the same metro contains several industries that do not share a narrative, and most directories force them into one. Defense and intelligence programs that buy systems, not slides. Integrators who live on delivery rather than on a Series B announcement. Construction firms and owners’ reps who live on specifications, drawings, and change orders. University and federal-lab spillover from Georgetown, Mason, Maryland, and Hopkins that does not always incorporate in Delaware and move to San Francisco. And small studios like ours that keep the GPU rack and the machine shop in the same 4,500 square feet, because that is how you get from a prototype to a thing you can hand somebody.
Joshua8.AI sits in that last bucket, in McLean, with a second door in Rehoboth Beach. TeraContext.AI, the company we are building out of it, sits in the construction bucket. Neither is a satellite of someone else’s ecosystem. The customers, the documents, the procurement clock, and the failure modes are all local. A paraphrase that drops a bidder’s exclusions clause does not show up as a leaderboard regression. It shows up as a bad bid. Oluwadi got that detail exactly right, including the part of our follow-up work we have not formally published: on construction statements of work, a model can hold a perfectly respectable similarity score while omitting the exclusions in most of the chains. That is a DMV-shaped problem even when the weights were trained in California. You cannot understand a GC-facing document product by averaging it with a consumer chatbot and a foundation-model lab, and you cannot understand a Reston diligence shop, a McLean studio, and a Bethesda health-system vendor by calling them all East Coast tech.
Two caveats, because directories lie in both directions
First, “DMV” can become its own cartoon, where every company is secretly a cleared prime and every founder is one meeting from the Pentagon. Oluwadi’s own edition intro leans that way; it calls the intelligence and defense agencies “the region’s actual economic anchor, not a side industry.” Fair as history. But most of the operating companies in this region are ordinary firms selling ordinary things to hospitals, contractors, and owners, and the edition only works if the ordinary firms stay in the picture. Defense is the anchor. It is not the harbor.
Second, a regional graph is only as good as its nodes. Catch-all categories, missing headquarters, and “AI” as a bucket will recreate the national mush at metro scale. Oluwadi’s changelog treats that as a maintenance problem rather than a launch feature, which is the right instinct: universities got their own tier in June, before the site went public, and an August pass on category precision cut untyped companies from 29 percent to 12.5 percent. They are also candid that classification and drafting are AI-assisted, with a model’s category suggestion sitting in a review queue before it changes anything. I have some professional interest in that design. Thirty rounds of paraphrase taught us that the errors which matter are the ones that keep a high score while the content quietly leaves. A directory that gets edited is a map. A directory that does not is a press list.
What I actually want from a regional edition
Not a trophy page. I want the edges a builder actually uses: who writes the specs, who owns the drawings, which investors will look at a construction workflow, which university labs still touch hardware, which two firms sit in the same county and never appear in the same article. If the graph can hold those without flattening them into “AI / software / semiconductors,” it is doing something the national lists cannot, and it is exactly the kind of map I would pay Osparna for.
They wrote us up as a technical result and a product rule. The larger courtesy is the shelf. This market is not a suburb. Treat it like one and you will miss the clause that changes the bid, and the company that was built to catch it.
Oluwadi’s directory: oluwadi.com. The DMV edition: Defense · Intelligence · Technology Capital. Their write-up of Joshua8.AI and TeraContext.AI is here.