Halluminate raises $30M Series A to train AI on real financial deal work

Fortune reported Thursday that San Francisco startup Halluminate raised a $30 million Series A led by Oak HC/FT, bringing total funding to $38.5 million, as it builds simulated financial workflows that show where frontier agents break on multi-day deal work.

Frontier models still stumble on financial work that runs for days, when the instructions change and the data room is a mess of files. Halluminate’s bet is that the next step is not another general chatbot, but practice environments built for that deal work — reinforcement learning that turns those failures into something a model can train on — and that the buyers paying for it today are the labs themselves.

On Thursday, 1 October 2026, Fortune reported that Halluminate raised $30 million in a Series A led by Oak HC/FT, bringing total funding to $38.5 million. A Series A is an early private funding round, named with a letter. Thirty million dollars is the new money. Thirty-eight and a half million is every dollar raised so far, including this round. Take $30 million off $38.5 million and the earlier raises add up to $8.5 million. That subtraction is arithmetic. Fortune does not name the earlier round. The piece is a Fortune exclusive by Wen Shao. The page dates it October 1, 2026, 11:00 a.m. Eastern. Halluminate is a nine-person startup in San Francisco that builds AI training environments for financial work. Those lines are Fortune’s.

What the company does, as Fortune states it. Halluminate benchmarks AI models on financial tasks to find where they fall short, and it builds simulated training environments aimed at those gaps. A benchmark is a scored test. A simulated environment is a practice version of the work, not the live deal. Founded in 2024, the company is betting that the data and the environments used to train AI will get more specialized by industry. Jerry Wu, the chief executive, told Fortune that simulating an investment banker’s work is fundamentally different from simulating a software engineer’s work. He calls the systems “verticalized data research labs.” Verticalized, in that phrase, means built for one industry’s work, not for every job at once. He expects training data to specialize, with companies going deep on finance, coding, or healthcare rather than across many fields. Those lines are Fortune’s, including Wu’s.

The August benchmark, as Fortune reports it. The company asked seven frontier models to work through a simulated company-acquisition due-diligence process. Frontier, here, means the leading general models. Due diligence is the check a buyer runs before it acquires a company: the contracts, the numbers, the risks. The test had 88 tasks, based on anonymized private-equity transactions and written and reviewed by practicing deal professionals. Anonymized means the real deal was stripped of names. Private equity is the business of buying companies with private money. The highest average score was 51 percent. That is a bit more than half, and it was the best average among the seven. Fortune does not name the model that scored it, and it does not print the other six averages. Those lines are Fortune’s.

One task inside that test, as Fortune describes it. An agent had to redline a statement of work using a 160-file data room, 21 emails across nine threads, and four meeting notes. An agent, here, is software that takes the next step in the work, not only a chat reply. A statement of work is the document that says what will be done, and on what terms. To redline it is to mark the changes. A data room is the folder of files a buyer is given. One hundred sixty files is a large pile for one task. As the deal terms changed, the agent had to find the latest instructions and still keep the provisions that were supposed to stay. Across the benchmark, Fortune says, the agents struggled to carry instructions through to the final deliverable. They left out required changes, used the wrong analytical method, or relied on information that had been replaced. Those lines are Fortune’s.

What the benchmark is for, as Fortune states it. Tests like this would show where agents break on complex financial workflows, then turn those failure modes into reinforcement-learning environments. A failure mode is the specific way the agent breaks. Reinforcement learning, here, is practice with a score: the model tries the task, gets told what it got wrong, and is trained on that. Those lines are Fortune’s.

Why Oak wrote the check, as Fortune reports it. Matt Streisfeld, a general partner at Oak HC/FT, told Fortune that finance offers a broad range of complex knowledge work, from banking and private equity to consulting and accounting. As agents take on work that stretches from hours into days, he expects the quality of specialized training environments to matter more. He said Halluminate’s finance expertise, and its setup for building high-quality environments, stood out. Quoted: “When the agent starts getting into long horizon work, testing work and specialization will really be key.” Long horizon, in that sentence, means a task that runs for a long stretch, not a single reply. That quotation is his, in Fortune.

The market around the round, as Fortune frames it. Demand for this kind of post-training infrastructure is showing up in customer projects and in deals. Post-training is the work after a general model exists: the extra practice that aims it at a job. Scale AI wrote in February that nearly half of its new data-training projects involve reinforcement-learning environments. Nearly half is a bit under 50 percent. Fortune says February and does not print the year beside it. Deeptune, which builds simulated work environments for training agents, raised a $43 million Series A led by Andreessen Horowitz in March and agreed to be acquired by Mercor four months later. Forty-three million is that other company’s round, not Halluminate’s. Fortune does not print a price for the Mercor deal. Those lines are Fortune’s context. They are not Halluminate’s results.

Customers and revenue, as Wu told Fortune. Four of the top five closed-source U.S. AI labs are paying customers. Closed-source means the lab does not publish the model’s weights for anyone to copy. Four of five is Wu’s count. Fortune does not name the four labs, and it does not name the one that is not a customer. He also said Halluminate has crossed the mid-eight figures in annualized revenue run rate, based on quarterly revenue from work already delivered and paid for, and that the company is profitable. Annualized means a yearly pace taken from a shorter period. A run rate is that pace, not cash already collected for a full year. Mid-eight figures means a yearly pace in the tens of millions. Eight figures runs from $10 million to just under $100 million. Fortune does not print the exact dollar figure. Profitable, in his sentence, means the company says it makes money, not only that it is growing. Those lines are Wu’s, in Fortune. Fortune does not print a valuation, the price a round puts on the whole company.

Where Wu says the company will aim, as Fortune reports him. For now, Halluminate is concentrating on a small group of frontier model labs, rather than selling broadly to companies that would use the tools for their own staff. He called that focus a strength. He wants to work first with the labs, where Halluminate can push what the models can do and develop how it builds the training environments. Enterprise customers and other industries could come later. Neither is a priority today. Enterprise, here, means a company buying the tool for its own people, rather than a lab training a model. Halluminate argues that specializing lets it compound finance expertise, its network of experts, and its methods for checking work and generating data. As models improve, the environments have to get harder or they stop being useful for training. Wu calls that pressure the “Moore’s law of environments.” Moore’s law, in chip history, is the rough rule that chips get much more capable on a steady clock. His version is an estimate: every six to eight months, the complexity needs to roughly double to keep pushing frontier models. Six to eight months is about two quarters to three. Roughly double means about twice as hard. He said that complexity can mean longer trajectories, harder reasoning, and more files. A trajectory is the path of steps the agent takes. He described the company’s intellectual property as the ability to keep producing that complexity “generation after generation.” Those lines are Fortune’s, including Wu’s. The six-to-eight-month estimate is his.

Who is in the round. The two accounts agree on the lead, the new money, and the total, and they do not list the same participants. Fortune, citing Wu, says existing investors Y Combinator, Orange Collective, and Heavybit participated, along with individual researchers from Anthropic, OpenAI, and Meta. Individual researchers means people who work at those labs, not a check from the labs themselves. Fortune does not say Anthropic, OpenAI, or Meta invested as companies. The same day, Wu posted a Series A note on Halluminate’s site. That note says the $30 million round was led by Oak HC/FT, with participation from Y Combinator, Orange Collective, FT Partners, and Heavybit, and that total capital raised is $38.5 million. FT Partners is on the company’s list. Fortune’s account does not name FT Partners. The company’s note does not name the individual researchers. Both accounts are dated 1 October 2026.

What the company site adds. The homepage calls Halluminate a data research lab that builds benchmarks and reinforcement-learning environments for knowledge work. It says the first focus is valuable workflows in financial services and consulting, and that the company is rapidly expanding into neighboring fields. Fortune quotes Wu saying other industries are not a priority today. The homepage’s expansion line and that quotation are both on the record. They are not the same sentence. The Series A note repeats the expansion line. It also gives a company scoreboard for the last 10 months: from zero to a mid-eight-figure revenue run rate while remaining strongly profitable; reinforcement-learning environments with four of the five leading closed-source U.S. AI labs; a due-diligence benchmark it calls Westworld; and a team that includes former founders, researchers, particle physicists, experts, and engineers from Meta, Scale AI, Capital One Labs, McKinsey, and Goldman Sachs. Strongly profitable is the company’s wording. Fortune’s wording, citing Wu, is profitable. Four of five leading labs, on the company note, is the same count Wu gave Fortune as paying customers, in slightly different words. The team list is the company’s. Fortune does not print it. The note is by Jerry Wu and dated October 1, 2026. The research page names the August test Westworld Finance Diligence Bench and dates that write-up 11 August 2026. It says the bench has 88 problems drawn from anonymized real private transactions, written and reviewed by practicing finance deal professionals, run inside desktop environments, on paths that can reach hundreds of steps. Fortune says August, 88 tasks, anonymized private-equity transactions, and practicing deal professionals, and it reports the highest average score as 51 percent. Fortune does not print the product name or the day in August. The company pages describe the bench and do not print that 51 percent.

The picture is a photograph of Halluminate’s founders on a Y Combinator demo-day stage. Two people in matching dark shirts stand at a wooden podium, one of them at the microphone. An orange backdrop carries the Y Combinator mark, and a large screen sits behind them. The slide is titled Halluminate. Its line reads “Data and RL environments to automate knowledge work,” with a second card about training models for real-world computer use. The Y Combinator mark also sits at the lower right of the frame. It is a stage photograph. It does not print a calendar date.

In plain terms, Fortune reported on Thursday that Halluminate, a nine-person San Francisco company founded in 2024, raised a $30 million Series A led by Oak HC/FT, and that total funding is $38.5 million. The company builds practice environments that show where frontier agents break on financial work that runs for days, then turns those breaks into training. In an August test of a simulated acquisition, seven models were scored on 88 tasks from anonymized private-equity deals, and the best average was 51 percent. Wu told Fortune that four of the top five closed-source U.S. AI labs are paying customers, and that the company has passed a mid-eight-figure yearly revenue pace and is profitable. Fortune did not name the labs, did not print the exact revenue figure, and did not print a valuation. The company’s own note the same day names FT Partners among the participants. Fortune’s account, citing Wu, names individual researchers from Anthropic, OpenAI, and Meta, and does not name FT Partners.